{"id":882,"date":"2023-12-10T19:43:43","date_gmt":"2023-12-10T16:43:43","guid":{"rendered":"https:\/\/www.facadium.com.tr\/blog\/?p=882"},"modified":"2023-12-11T00:20:22","modified_gmt":"2023-12-10T21:20:22","slug":"bilgisayarli-tibbi-goruntuleme-ve-grafik","status":"publish","type":"post","link":"https:\/\/www.facadium.com.tr\/blog\/bilgisayarli-tibbi-goruntuleme-ve-grafik\/","title":{"rendered":"B\u0130LG\u0130SAYARLI TIBB\u0130 G\u00d6R\u00dcNT\u00dcLEME VE GRAF\u0130K"},"content":{"rendered":"\n<h1 class=\"wp-block-heading has-text-align-center\" id=\"imagecas-bilgisayarli-tomografi-anjiyografi-goruntulerine-dayali-koroner-arter-segmentasyonu-icin-genis-olcekli-bir-veri-seti-ve-referans-noktasi-calismasi\">IMAGECAS: B\u0130LG\u0130SAYARLI TOMOGRAF\u0130 ANJ\u0130YOGRAF\u0130 G\u00d6R\u00dcNT\u00dcLER\u0130NE DAYALI KORONER ARTER SEGMENTASYONU \u0130\u00c7\u0130N GEN\u0130\u015e \u00d6L\u00c7EKL\u0130 B\u0130R VER\u0130 SET\u0130 VE REFERANS NOKTASI \u00c7ALI\u015eMASI<\/h1>\n\n\n\n<p><strong>Not : <\/strong>Bilgisayarl\u0131 T\u0131bbi G\u00f6r\u00fcnt\u00fcleme Ve Grafik \u00fczerine yap\u0131lan, Imagecas: Bilgisayarl\u0131 Tomografi Anjiyografi G\u00f6r\u00fcnt\u00fclerine Dayal\u0131 Koroner Arter Segmentasyonu \u0130\u00e7in Geni\u015f \u00d6l\u00e7ekli Bir Veri Seti Ve Referans Noktas\u0131 adl\u0131 \u00e7al\u0131\u015fma, bilimsel \u00e7al\u0131\u015fmalara katk\u0131 sa\u011flamak ve T\u00fcrk literat\u00fcr\u00fcne girmesi ad\u0131na Xiaowei Xu&#8217;nun vermi\u015f oldu\u011fu izinle, taraf\u0131m\u0131zca T\u00fcrk\u00e7e&#8217;ye \u00e7evrilmi\u015f ve yay\u0131nlanm\u0131\u015ft\u0131r.<\/p>\n\n\n\n<div class=\"wp-block-rank-math-toc-block\" id=\"rank-math-toc\"><h2>\u0130\u00e7indekiler<\/h2><nav><ul><li class=\"\"><a href=\"#imagecas-bilgisayarli-tomografi-anjiyografi-goruntulerine-dayali-koroner-arter-segmentasyonu-icin-genis-olcekli-bir-veri-seti-ve-referans-noktasi-calismasi\">IMAGECAS: B\u0130LG\u0130SAYARLI TOMOGRAF\u0130 ANJ\u0130YOGRAF\u0130 G\u00d6R\u00dcNT\u00dcLER\u0130NE DAYALI KORONER ARTER SEGMENTASYONU \u0130\u00c7\u0130N GEN\u0130\u015e \u00d6L\u00c7EKL\u0130 B\u0130R VER\u0130 SET\u0130 VE REFERANS NOKTASI \u00c7ALI\u015eMASI<\/a><ul><li class=\"\"><a href=\"#ozet\">\u00d6ZET<\/a><\/li><li class=\"\"><a href=\"#giris\">Giri\u015f<\/a><\/li><li class=\"\"><a href=\"#ilgili-calisma\">\u0130lgili \u00c7al\u0131\u015fma<\/a><ul><li class=\"\"><a href=\"#geleneksel-ml-tabanli-yaklasim\">Geleneksel ML Tabanl\u0131 Yakla\u015f\u0131m<\/a><\/li><li class=\"\"><a href=\"#dl-tabanli-yaklasim\">DL-Tabanl\u0131 Yakla\u015f\u0131m<\/a><\/li><li class=\"\"><a href=\"#karsilastirma-ve-veri-kumeleri\">Kar\u015f\u0131la\u015ft\u0131rma ve Veri K\u00fcmeleri<\/a><\/li><\/ul><\/li><li class=\"\"><a href=\"#image-cad-veriseti\">ImageCAD Veriseti<\/a><\/li><li class=\"\"><a href=\"#karsilastirma\">Kar\u015f\u0131la\u015ft\u0131rma<\/a><ul><li class=\"\"><a href=\"#dogrudan-segmentasyon\">Do\u011frudan Segmentasyon<\/a><\/li><\/ul><\/li><li class=\"\"><a href=\"#yama-tabanli-segmentasyon\">Yama Tabanl\u0131 Segmentasyon<\/a><\/li><li class=\"\"><a href=\"#agac-verilerine-dayali-segmentasyon\">A\u011fa\u00e7 Verilerine Dayal\u0131 Segmentasyon<\/a><\/li><li class=\"\"><a href=\"#grafik-tabanli-segmentasyon\">Grafik Tabanl\u0131 Segmentasyon<\/a><\/li><li class=\"\"><a href=\"#temel-yontem\">Temel Y\u00f6ntem<\/a><\/li><li class=\"\"><a href=\"#deneyler-ve-tartisma\">Deneyler ve Tart\u0131\u015fma<\/a><\/li><li class=\"\"><a href=\"#deneme-kurulumu\">Deneme Kurulumu<\/a><\/li><li class=\"\"><a href=\"#yapilandirma-tartismasi\">Yap\u0131land\u0131rma Tart\u0131\u015fmas\u0131<\/a><ul><li class=\"\"><a href=\"#dogrudan-segmentasyon-1\">Do\u011frudan Segmentasyon<\/a><\/li><li class=\"\"><a href=\"#yama-tabanli-segmentasyon-2\">Yama Tabanl\u0131 Segmentasyon<\/a><\/li><li class=\"\"><a href=\"#agac-verisine-dayali-segmentasyon-ve-grafik-tabanli-segmentasyon\">A\u011fa\u00e7 Verisine Dayal\u0131 Segmentasyon ve Grafik Tabanl\u0131 Segmentasyon<\/a><\/li><li class=\"\"><a href=\"#temel-yontem-3\">Temel Y\u00f6ntem<\/a><\/li><\/ul><\/li><li class=\"\"><a href=\"#olcut-karsilastirmasi\">\u00d6l\u00e7\u00fct Kar\u015f\u0131la\u015ft\u0131rmas\u0131<\/a><\/li><li class=\"\"><a href=\"#tartisma\">Tart\u0131\u015fma<\/a><\/li><\/ul><\/li><li class=\"\"><a href=\"#cozum\">\u00c7\u00f6z\u00fcm<\/a><\/li><li class=\"\"><a href=\"#yazar-katki-beyani\">Yazar Katk\u0131 Beyan\u0131<\/a><\/li><li class=\"\"><a href=\"#rekabetci-menfaat-beyani\">Rekabet\u00e7i Menfaat Beyan\u0131<\/a><\/li><li class=\"\"><a href=\"#veri-kullanilabilirligi\">Veri Kullan\u0131labilirli\u011fi<\/a><\/li><li class=\"\"><a href=\"#tesekkurler\">Te\u015fekk\u00fcrler<\/a><\/li><li class=\"\"><a href=\"#etik-ve-bilgi-yonetimi-onaylari\">Etik ve Bilgi Y\u00f6netimi Onaylar\u0131<\/a><\/li><li class=\"\"><a href=\"#referanslar\">Referanslar<\/a><\/li><\/ul><\/nav><\/div>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"ozet\">\u00d6ZET<\/h2>\n\n\n\n<p>     Kardiyovask\u00fcler hastal\u0131klar (KVH), bula\u015f\u0131c\u0131 olmayan hastal\u0131klar\u0131n yakla\u015f\u0131k yar\u0131s\u0131n\u0131 olu\u015fturur. Koroner arterdeki damar stenozu, KVH&#8217;n\u0131n en b\u00fcy\u00fck riski olarak kabul edilmektedir. Bilgisayarl\u0131 tomografi anjiyografi (BTA), \u00fcst\u00fcn g\u00f6r\u00fcnt\u00fc \u00e7\u00f6z\u00fcn\u00fcrl\u00fc\u011f\u00fc nedeniyle koroner arter tan\u0131s\u0131nda yayg\u0131n olarak kullan\u0131lan noninvaziv g\u00f6r\u00fcnt\u00fcleme y\u00f6ntemlerinden biridir. Klinik olarak koroner arter hastal\u0131\u011f\u0131n\u0131n tan\u0131s\u0131 ve miktar\u0131n\u0131n belirlenmesi i\u00e7in koroner arterlerin segmentasyonu \u00f6nemlidir. Son zamanlarda bu soruna \u00e7\u00f6z\u00fcm bulmak i\u00e7in \u00e7e\u015fitli \u00e7al\u0131\u015fmalar \u00f6nerilmi\u015ftir.<\/p>\n\n\n\n<p>     Bununla birlikte, bir yandan \u00e7o\u011fu \u00e7al\u0131\u015fma \u015firket i\u00e7i veri k\u00fcmelerine dayan\u0131yor ve yaln\u0131zca birka\u00e7 \u00e7al\u0131\u015fma, yaln\u0131zca onlarca g\u00f6r\u00fcnt\u00fc i\u00e7eren veri k\u00fcmelerini kamuya yay\u0131nl\u0131yor. \u00d6te yandan, kaynak kodlar\u0131 yay\u0131nlanmam\u0131\u015ft\u0131r ve takip \u00e7al\u0131\u015fmalar\u0131n\u0131n \u00e7o\u011fu mevcut \u00e7al\u0131\u015fmalarla kar\u015f\u0131la\u015ft\u0131rma yapmam\u0131\u015ft\u0131r, bu da y\u00f6ntemlerin etkilili\u011fini de\u011ferlendirmeyi zorla\u015ft\u0131rmakta ve bu zorlu ama kritik sorunun daha fazla ara\u015ft\u0131r\u0131lmas\u0131n\u0131 engellemektedir. <\/p>\n\n\n\n<p>     Bu yaz\u0131da, CTA g\u00f6r\u00fcnt\u00fclerinde koroner arter segmentasyonu i\u00e7in geni\u015f \u00f6l\u00e7ekli bir veri seti \u00f6neriyoruz. Ek olarak, mevcut birka\u00e7 tipik y\u00f6ntemi uygulamak i\u00e7in elimizden gelenin en iyisini yapmaya \u00e7al\u0131\u015ft\u0131\u011f\u0131m\u0131z bir k\u0131yaslama uygulad\u0131k. Ayr\u0131ca, damarlar\u0131n ayr\u0131nt\u0131lar\u0131n\u0131 \u00e7\u0131karmak i\u00e7in \u00e7ok \u00f6l\u00e7ekli yama f\u00fczyonunu ve iki a\u015famal\u0131 i\u015flemeyi birle\u015ftiren g\u00fc\u00e7l\u00fc bir temel y\u00f6ntem \u00f6neriyoruz. Kapsaml\u0131 deneyler, \u00f6nerilen y\u00f6ntemin, \u00f6nerilen b\u00fcy\u00fck \u00f6l\u00e7ekli veri seti \u00fczerinde mevcut \u00e7al\u0131\u015fmalardan daha iyi performans elde etti\u011fini g\u00f6stermektedir.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"giris\">Giri\u015f<\/h2>\n\n\n\n<p>     Kardiyovask\u00fcler hastal\u0131k (KVH) g\u00fcn\u00fcm\u00fczde d\u00fcnyan\u0131n \u00f6nde gelen sa\u011fl\u0131k sorunlar\u0131ndan biridir. D\u00fcnya Sa\u011fl\u0131k \u00d6rg\u00fct\u00fc&#8217;ne (WHO) g\u00f6re, 2019 y\u0131l\u0131nda KVH&#8217;ye ba\u011fl\u0131 17,9 milyon \u00f6l\u00fcm meydana geldi ve bu, t\u00fcm k\u00fcresel \u00f6l\u00fcmlerin %32&#8217;sini olu\u015fturdu (Organizasyon ve ark., 2009). Avustralya Sa\u011fl\u0131k ve Refah Enstit\u00fcs\u00fc (AIHW), KVH&#8217;nin Avustralya&#8217;daki \u00f6l\u00fcmlerin \u00f6nde gelen nedeni oldu\u011funu ve 2018&#8217;deki t\u00fcm \u00f6l\u00fcmlerin %42&#8217;sini temsil etti\u011fini bildirdi (Zhang, 2010). T\u00fcm CVD&#8217;ler aras\u0131nda koroner kalp hastal\u0131\u011f\u0131, patofizyolojisinin temel olarak anormal koroner arter stenozu ile ili\u015fkilendirildi\u011fi en yayg\u0131n tiptir (Cooper ve ark., 2000). <\/p>\n\n\n\n<p>     Bu t\u00fcr stenoz s\u0131kl\u0131kla miyokardiyal perf\u00fczyonun azalmas\u0131na ve miyokardiyal h\u00fccrelerin hipoksi hasar\u0131na neden olur ve sonunda miyokard enfarkt\u00fcs\u00fcne yol a\u00e7ar. Klinik pratikte bilgisayarl\u0131 tomografi anjiyografi (BTA), noninvaziv olmas\u0131 ve y\u00fcksek \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc 3 boyutlu g\u00f6r\u00fcnt\u00fcleme sa\u011flayabilmesi nedeniyle koroner arter hastal\u0131klar\u0131n\u0131n tan\u0131 ve tedavi planlamas\u0131nda yayg\u0131n olarak kullan\u0131lmaktad\u0131r (Collet ve ark., 2018). BTA g\u00f6r\u00fcnt\u00fcleri al\u0131nd\u0131ktan sonra radyologlar \u00f6ncelikle koroner arterlerin yerini tespit eder ve s\u0131n\u0131rlar\u0131n\u0131 izole eder. Daha sonra nihai te\u015fhis ve tedavi planlamas\u0131 i\u00e7in daralan k\u0131s\u0131m \u00e7\u0131kart\u0131larak miktar\u0131 belirlenir.<\/p>\n\n\n\n<p>     Bununla birlikte, man\u00fcel i\u015flemlere dayanan bu t\u00fcr bir prosesin, olumsuz bir \u015fekilde zaman al\u0131c\u0131 ve hataya a\u00e7\u0131k oldu\u011fu yayg\u0131n olarak kabul edilmektedir. Daha da k\u00f6t\u00fcs\u00fc, t\u0131bbi g\u00f6r\u00fcnt\u00fclerin s\u00fcrekli artan miktar\u0131 ve \u00e7e\u015fitlili\u011fi (Li ve di\u011ferleri, 2018c), maliyet ve tekrarlanabilirlik a\u00e7\u0131s\u0131ndan manuel segmentasyonu tamamen uygulanamaz hale getirebilir. <\/p>\n\n\n\n<p>     Bu nedenle otomatik koroner arter segmentasyonu olduk\u00e7a arzu edilir. Ancak bu g\u00f6rev bir\u00e7ok nedenden dolay\u0131 olduk\u00e7a zordur. <\/p>\n\n\n\n<p>     Birincisi, koroner arterin anatomik yap\u0131s\u0131 pop\u00fclasyondan pop\u00fclasyona \u00f6nemli \u00f6l\u00e7\u00fcde de\u011fi\u015fmektedir. \u00d6rne\u011fin, koroner arterler genellikle bir ya\u011f tabakas\u0131yla \u00e7evrilidir ancak baz\u0131 insanlarda arterler kalp kas\u0131n\u0131n i\u00e7indedir. <\/p>\n\n\n\n<p>     \u0130kincisi, baz\u0131 CTA g\u00f6r\u00fcnt\u00fcleri, segmentasyon sonu\u00e7lar\u0131nda d\u00fc\u015f\u00fck kaliteye neden olacak artefaktlar nedeniyle g\u00fcr\u00fclt\u00fcl\u00fcd\u00fcr. <\/p>\n\n\n\n<p>     \u00dc\u00e7\u00fcnc\u00fcs\u00fc, koroner arterlerin t\u00fcb\u00fcler yap\u0131s\u0131 son derece karma\u015f\u0131kt\u0131r. \u00d6rne\u011fin, arterler boyunca \u00e7ok say\u0131da \u00e7atallanma vard\u0131r ve enine d\u00fczlemlerde az miktarda koroner arter alan\u0131 vard\u0131r (Zhu ve ark., 2021). <\/p>\n\n\n\n<p>     Yukar\u0131daki konular\u0131 ele almak i\u00e7in son on y\u0131lda toplulukta onlarca \u00e7al\u0131\u015fma \u00f6nerilmi\u015ftir. Bu \u00e7al\u0131\u015fmalarda kullan\u0131lan yakla\u015f\u0131mlar iki ana kategoriye ayr\u0131labilir: geleneksel makine \u00f6\u011frenimi (ML) tabanl\u0131 y\u00f6ntem ve derin \u00f6\u011frenme (DL) tabanl\u0131 y\u00f6ntem. Geleneksel ML tabanl\u0131 y\u00f6ntem ayr\u0131ca piksel tabanl\u0131 y\u00f6ntem ve yap\u0131 tabanl\u0131 y\u00f6ntem olarak alt s\u0131n\u0131flara ayr\u0131labilir (Doyle et al., 2006; Nguyen et al., 2012; Tabesh et al.,2007; Sirinukunwattana et al., 2015a). <\/p>\n\n\n\n<p>     Bu y\u00f6ntemler, el yap\u0131m\u0131 \u00f6zellikleri ve koroner arter yap\u0131lar\u0131na ili\u015fkin \u00f6n bilgileri kullanarak umut verici sonu\u00e7lar elde etmektedir (Zheng ve di\u011ferleri, 2011; Mohr ve di\u011ferleri, 2012; Broersen ve di\u011ferleri, 2012; Shahzad ve di\u011ferleri, 2013; Chi ve di\u011ferleri). Ancak ciddi deformasyona sahip koronerlere uyguland\u0131\u011f\u0131nda \u00f6nemli \u00f6l\u00e7\u00fcde bozulmaya u\u011frarlar. ML tabanl\u0131 y\u00f6ntemden farkl\u0131 olarak, yak\u0131n zamanda \u00f6nerilen DL tabanl\u0131 y\u00f6ntemler, \u00e7ok az el yap\u0131m\u0131 \u00f6zellik veya \u00f6n bilgi gerektirir. <\/p>\n\n\n\n<p>     B\u00f6yle bir yakla\u015f\u0131m\u0131 kullanan \u00e7al\u0131\u015fmalar, ilk yakla\u015f\u0131ma g\u00f6re \u00f6nemli bir geli\u015fme elde etmi\u015f ve koroner arter segmentasyonu i\u00e7in y\u00fcksek etkinli\u011fini kan\u0131tlam\u0131\u015ft\u0131r (Huang et al., 2018; Shen et al., 2019; Chen et al., 2019; Wolterink et al., 2019; Kong et al., 2020; Gu and Cai, 2021; Zhu et al., 2021; Tian et al., 2021).<\/p>\n\n\n\n<p>     Mevcut \u00e7al\u0131\u015fmalar\u0131n Tablo 1 ve Tablo 2&#8217;de g\u00f6sterildi\u011fi gibi ayr\u0131nt\u0131l\u0131 bir analizini yap\u0131yoruz ve \u00e7o\u011fu \u00e7al\u0131\u015fman\u0131n di\u011ferleriyle adil ve kapsaml\u0131 bir kar\u015f\u0131la\u015ft\u0131rma yapamad\u0131\u011f\u0131n\u0131 g\u00f6r\u00fcyoruz. \u00d6rne\u011fin Kong ve ark. (2020) daha \u00f6nceki ilgili \u00e7al\u0131\u015fmalarla kar\u015f\u0131la\u015ft\u0131rma yapmam\u0131\u015f, yaln\u0131zca baz\u0131 pop\u00fcler derin sinir a\u011flar\u0131 (DNN&#8217;ler) ile kar\u015f\u0131la\u015ft\u0131rma yapm\u0131\u015ft\u0131r. Shen ve ark. (2019), kar\u015f\u0131la\u015ft\u0131rma i\u00e7in ilgili \u00e7al\u0131\u015fmalar\u0131n Zar puan\u0131n\u0131 listeledi; bu, iki y\u00f6ntem ayn\u0131 veri seti kullan\u0131larak de\u011ferlendirilmedi\u011finden yeterince adil de\u011fil. Hatta baz\u0131 \u00e7al\u0131\u015fmalarda (Chi et al., 2015; Han et al., 2016) belirli klinik ihtiya\u00e7lara g\u00f6re uyarlanm\u0131\u015f farkl\u0131 de\u011ferlendirme metrikleri setleri kullan\u0131l\u0131rken, di\u011ferleri (Shen et al., 2019) ba\u015flang\u0131\u00e7taki aortun dahil edildi\u011fi farkl\u0131 a\u00e7\u0131klamalar bile kullanm\u0131\u015f ve bu da \u00e7ok daha y\u00fcksek bir Zar puan\u0131na yol a\u00e7m\u0131\u015ft\u0131r. <\/p>\n\n\n\n<p>     Bu de\u011ferlendirme \u00f6nyarg\u0131lar\u0131 temel olarak kamuya a\u00e7\u0131k geni\u015f \u00f6l\u00e7ekli bir k\u0131yaslama veri setinin bulunmamas\u0131ndan kaynaklanmaktad\u0131r. Koroner arter segmentasyonuna y\u00f6nelik yaln\u0131zca iki genel veri k\u00fcmesi (Schaap et al., 2009a; Kiri\u015fli et al., 2013) (also shown in Table 2), e\u011fitim i\u00e7in s\u0131ras\u0131yla yaln\u0131zca 8 ve 18 g\u00f6r\u00fcnt\u00fc i\u00e7erir. Bunun yan\u0131 s\u0131ra, mevcut y\u00f6ntemlerin hepsinin kaynak kodu yay\u0131nlanmamas\u0131, adil bir kar\u015f\u0131la\u015ft\u0131rma i\u00e7in daha fazla zorluk ortaya \u00e7\u0131karm\u0131\u015ft\u0131r.<\/p>\n\n\n\n<p>     Bu yaz\u0131da, otomatik koroner arter segmentasyon y\u00f6ntemlerinin etkinli\u011fini adil bir \u015fekilde ara\u015ft\u0131rmak i\u00e7in geni\u015f bir veri seti \u00f6neriyoruz. Bu veri k\u00fcmesi, mevcut genel veri k\u00fcmelerinden olduk\u00e7a b\u00fcy\u00fck olan 1000 3D CTA g\u00f6r\u00fcnt\u00fcs\u00fc i\u00e7erir. Ek olarak, bu veri setine dayal\u0131 olarak, yaln\u0131zca birka\u00e7 mevcut tipik y\u00f6ntemi uygulamakla kalmay\u0131p ayn\u0131 zamanda g\u00fc\u00e7l\u00fc bir temel y\u00f6ntem de \u00f6nerdi\u011fimiz bir k\u0131yaslama noktas\u0131 da \u00f6neriyoruz. Deneysel sonu\u00e7lar, temel y\u00f6ntemimizin mevcut t\u00fcm y\u00f6ntemlerden daha iyi performansa ula\u015ft\u0131\u011f\u0131n\u0131 g\u00f6stermektedir. Ve bu arada iyile\u015ftirme i\u00e7in iyi bir potansiyel g\u00f6steriyor. Bu \u00e7al\u0131\u015fman\u0131n katk\u0131lar\u0131 \u015fu \u015fekilde \u00f6zetlenebilir:<\/p>\n\n\n\n<p>\u2022 Koroner arter segmentasyonu i\u00e7in 1000 hastay\u0131 i\u00e7eren geni\u015f \u00f6l\u00e7ekli, kamuya a\u00e7\u0131k bir veri seti toplad\u0131k. Veri k\u00fcmesinin resmi bir veri b\u00f6l\u00fcm\u00fc de sa\u011flanmaktad\u0131r. Bu veri k\u00fcmesinin toplulukta ilgili ara\u015ft\u0131rmalar\u0131n desteklenmesine yard\u0131mc\u0131 olabilece\u011fini umuyoruz;<\/p>\n\n\n\n<p><br>\u2022 Koroner arter segmentasyonu i\u00e7in mevcut \u00e7e\u015fitli y\u00f6ntemleri uygulad\u0131\u011f\u0131m\u0131z bir k\u0131yaslama \u00f6nerdik. Kar\u015f\u0131la\u015ft\u0131rmal\u0131 de\u011ferlendirmeyi de yay\u0131nlad\u0131k ve bunun daha sonraki \u00e7al\u0131\u015fmalarda adil kar\u015f\u0131la\u015ft\u0131rmalar yap\u0131lmas\u0131na yard\u0131mc\u0131 olabilece\u011fini umuyoruz;<\/p>\n\n\n\n<p><br>\u2022 Damarlar\u0131n ayr\u0131nt\u0131lar\u0131n\u0131 \u00e7\u0131karmak i\u00e7in \u00e7ok \u00f6l\u00e7ekli yama f\u00fczyonunu ve iki a\u015famal\u0131 i\u015flemeyi birle\u015ftiren g\u00fc\u00e7l\u00fc bir temel y\u00f6ntem \u00f6nerdik ve deneysel sonu\u00e7lar, y\u00f6ntemimizin mevcut en geli\u015fmi\u015f y\u00f6ntemlerden daha iyi performans g\u00f6sterdi\u011fini g\u00f6steriyor.<\/p>\n\n\n\n<p>     Makalenin geri kalan\u0131 \u015fu \u015fekilde organize edilmi\u015ftir. B\u00f6l\u00fcm 2 ilgili \u00e7al\u0131\u015fmalara genel bir bak\u0131\u015f sunmaktad\u0131r. B\u00f6l\u00fcm 3&#8217;te \u00f6nerilen veri setinin detaylar\u0131 sunulmaktad\u0131r. Daha sonra, B\u00f6l\u00fcm 4&#8217;te mevcut birka\u00e7 tipik y\u00f6ntemi ve temel y\u00f6ntemimizi i\u00e7eren \u00f6nerilen k\u0131yaslama sunulmaktad\u0131r. Deney sonu\u00e7lar\u0131 B\u00f6l\u00fcm 5&#8217;te sunulmakta ve tart\u0131\u015f\u0131lmaktad\u0131r ve B\u00f6l\u00fcm 6&#8217;da makale sonu\u00e7lanmaktad\u0131r.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"ilgili-calisma\">\u0130lgili \u00c7al\u0131\u015fma<\/h2>\n\n\n\n<p>     Bu b\u00f6l\u00fcmde koroner arter segmentasyonu ile ilgili mevcut literat\u00fcr\u00fc g\u00f6zden ge\u00e7irece\u011fiz. \u0130lgili \u00e7al\u0131\u015fmalar\u0131 \u00f6ncelikle kullan\u0131lan yakla\u015f\u0131m t\u00fcrlerine g\u00f6re (geleneksel ML tabanl\u0131 yakla\u015f\u0131m ve DL tabanl\u0131 yakla\u015f\u0131m) inceliyoruz ve ard\u0131ndan bu \u00e7al\u0131\u015fmalarda yayg\u0131n olarak benimsenen veri k\u00fcmelerini ve kriterleri tart\u0131\u015f\u0131yoruz.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"geleneksel-ml-tabanli-yaklasim\">Geleneksel ML Tabanl\u0131 Yakla\u015f\u0131m<\/h3>\n\n\n\n<p>     El yap\u0131m\u0131 \u00f6zellikler ve koroner arter yap\u0131lar\u0131na ili\u015fkin \u00f6n bilgiler, Table 1. Zheng et al. (2011) Zheng&#8217;in, uzman a\u00e7\u0131klamal\u0131 bir veri k\u00fcmesine yerle\u015ftirilmi\u015f zengin alana \u00f6zg\u00fc bilgiden (\u00f6zellikle bir dizi geometrik ve g\u00f6r\u00fcnt\u00fc \u00f6zelli\u011finden) yararlanmak i\u00e7in bir ML y\u00f6ntemi \u00f6nerdi\u011fi gibi, bu yakla\u015f\u0131mda yayg\u0131n olarak kullan\u0131lmaktad\u0131r. Mohr ve ark. (2012), verimli i\u015fleme i\u00e7in seviye setine dayal\u0131 bir yakla\u015f\u0131m \u00f6nerdi. Wang ve ark. (2012), do\u011fru segmentasyon i\u00e7in seviye k\u00fcmelerini damarlar\u0131n \u00f6rt\u00fcl\u00fc 3 boyutlu modeliyle birle\u015ftiriyor. Broersen ve ark. (2012) birbirini takip eden \u00fc\u00e7 ad\u0131mdan olu\u015fan bir boru hatt\u0131n\u0131 benimsedi. Shahzad ve ark. (2013), \u00e7\u0131kar\u0131lan merkez \u00e7izgilerinin yard\u0131m\u0131yla segmentasyonu ger\u00e7ekle\u015ftirdi.<\/p>\n\n\n\n<p>     Lugauer ve ark. (2014a), yo\u011fun \u0131\u015f\u0131n d\u00f6k\u00fcm\u00fc yoluyla sa\u011flam bir l\u00fcmen kontur alg\u0131lamas\u0131n\u0131 m\u00fcmk\u00fcn k\u0131lmak i\u00e7in \u00f6\u011frenmeye dayal\u0131 bir s\u0131n\u0131r dedekt\u00f6r\u00fc kulland\u0131. Lugauer, d\u0131\u015fb\u00fckey \u00f6nceliklere sahip Markov rastgele alan form\u00fclasyonuna dayanan model k\u0131lavuzlu bir b\u00f6l\u00fcmleme yakla\u015f\u0131m\u0131 \u00f6nerdi. Chi ve ark. (2015), Yo\u011funluk, yerel \u015fekil ve k\u00fcresel yap\u0131n\u0131n koroner arter \u00f6zelliklerini bir \u00f6\u011frenme \u00e7er\u00e7evesine entegre etti. Lesage ve ark. (2016), damar b\u00f6l\u00fcmlendirmesini yinelemeli bir izleme s\u00fcreci olarak de\u011ferlendirdi ve koroner arterlerin tan\u0131mlanmas\u0131 i\u00e7in par\u00e7ac\u0131k filtrelerine dayal\u0131 yeni bir Bayesian izleme algoritmas\u0131 \u00f6nerdi. <\/p>\n\n\n\n<p>     Han ve ark. (2016), dallar\u0131 ve g\u00f6r\u00fcn\u00fc\u015fte ba\u011flant\u0131s\u0131z ama asl\u0131nda ba\u011flant\u0131l\u0131 damar b\u00f6l\u00fcmlerini bulmak i\u00e7in aktif bir arama y\u00f6ntemi kulland\u0131. Freiman, koroner a\u011fa\u00e7larda (Nickisch ve ark., 2015) k\u0131smi hacim etkilerini hesaba katan bir ak\u0131\u015f sim\u00fclasyon y\u00f6ntemi kulland\u0131 (Glover ve Pelc, 1980). Gao ve ark. (2019), dairesel Hough d\u00f6n\u00fc\u015f\u00fcm\u00fc kullanarak aortu \u00e7\u0131kararak koroner k\u00f6k\u00fcn yerini tespit etti. Du ve ark (2021), g\u00fcr\u00fclt\u00fc azaltma, aday b\u00f6lge tespiti, geometrik \u00f6zellik \u00e7\u0131karma ve koroner arter izleme tekniklerini i\u00e7eren yeni bir segmentasyon \u00e7er\u00e7evesi \u00f6nerdi.<\/p>\n\n\n\n<p><strong>Tablo 1<\/strong><\/p>\n\n\n\n<p>     Mevcut geleneksel makine \u00f6\u011frenimi tabanl\u0131 \u00e7al\u0131\u015fmalar ve son on y\u0131ldaki koroner arter segmentasyonunun Zar puanlar\u0131 (% olarak). \ud835\udc4e(\ud835\udc4f)&#8217;da \ud835\udc4e genel miktard\u0131r ve \ud835\udc4f e\u011fitim miktar\u0131d\u0131r. Genel, y\u00f6ntemin \u00e7e\u015fitli uygulamalar i\u00e7in tasarland\u0131\u011f\u0131n\u0131 belirtirken, \u00f6zel, y\u00f6ntemin koroner damar segmentasyonu i\u00e7in tasarland\u0131\u011f\u0131n\u0131 ve optimize edildi\u011fini belirtir.<\/p>\n\n\n\n<p>a Yaln\u0131zca test verileri s\u00f6z konusudur (e\u011fitim verileri ge\u00e7erli de\u011fildir).<br>b Segmentasyon sonucu aortun ba\u015flang\u0131\u00e7 k\u0131sm\u0131n\u0131 i\u00e7erir.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"853\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-1-1024x853.webp\" alt=\"Mevcut geleneksel makine \u00f6\u011frenimi tabanl\u0131 \u00e7al\u0131\u015fmalar ve son on y\u0131ldaki koroner arter segmentasyonunun Zar puanlar\u0131\" class=\"wp-image-883\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-1-1024x853.webp 1024w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-1-300x250.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-1-768x640.webp 768w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-1-jpg.webp 1030w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">Mevcut geleneksel makine \u00f6\u011frenimi tabanl\u0131 \u00e7al\u0131\u015fmalar ve son on y\u0131ldaki koroner arter segmentasyonunun Zar puanlar\u0131<\/figcaption><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"dl-tabanli-yaklasim\">DL-Tabanl\u0131 Yakla\u015f\u0131m<\/h3>\n\n\n\n<p>     Derin \u00f6\u011frenmenin y\u00fckseli\u015finden bu yana ilgili topluluklarda b\u00fcy\u00fck ilgi g\u00f6rd\u00fc. \u015eu anda, Tablo 2&#8217;de g\u00f6sterildi\u011fi gibi, piksel bazl\u0131 segmentasyon (Moeskops et al., 2016; Kjerland, 2017), do\u011frudan segmentasyon (Shen et al., 2019; Lee et al., 2019; Yang et al., 2019; Fu et al., 2020; Lei et al., 2020; Gu et al., 2020; Gu and Cai, 2021; Zhu et al., 2021; Liang et al., 2021; <\/p>\n\n\n\n<p>     Tian et al., 2021; Cheung et al., 2021; Li et al., 2018a,b; Lin et al., 2022), yama bazl\u0131 segmentasyon (Duan et al., 2018; Chen et al., 2018b; Huang et al., 2018; Chen et al., 2019; Mirunalini et al., 2019; Wang et al., 2021; Pan et al., 2021), a\u011fa\u00e7 veri bazl\u0131 segmentasyon (Kong et al., 2020) ve grafik veri bazl\u0131 segmentasyon (Wolterink et al., 2019) dahil olmak \u00fczere temel olarak be\u015f teknik e\u011filim bulunmaktad\u0131r. <\/p>\n\n\n\n<p>     Koroner arterlerin detayl\u0131 anatomik etiketlenmesi gibi di\u011fer \u00e7al\u0131\u015fmalar da bu makalenin kapsam\u0131 d\u0131\u015f\u0131ndad\u0131r.<\/p>\n\n\n\n<p>     Piksel tabanl\u0131 segmentasyon, evri\u015fimli sinir a\u011flar\u0131n\u0131 (CNN&#8217;ler) koroner arter segmentasyonuna uyarlaman\u0131n \u00f6nc\u00fcs\u00fcd\u00fcr. Moeskops ve ark. (2016), koroner damar segmentasyonu i\u00e7in derin \u00f6\u011frenmenin uygulanabilirli\u011fini g\u00f6stermek amac\u0131yla tek bir evri\u015fimli sinir a\u011f\u0131 kullanm\u0131\u015ft\u0131r. Kjerland (2017), s\u0131ras\u0131yla aort segmentasyonu ve koroner segmentasyonu konusunda e\u011fitilmi\u015f ve koroner arter a\u011fac\u0131n\u0131n tamam\u0131n\u0131 segmentlere ay\u0131rabilen iki sinir a\u011f\u0131 kullanm\u0131\u015ft\u0131r. <\/p>\n\n\n\n<p>     Daha sonra U net&#8217;in y\u00fckseli\u015fi nedeniyle do\u011frudan segmentasyon pop\u00fcler hale geldi (Ronneberger ve di\u011ferleri, 2015). Shen ve ark. (2019), derin \u00f6\u011frenmeye ve geleneksel seviye belirleme y\u00f6ntemine dayal\u0131 ortak bir \u00e7er\u00e7eve \u00e7al\u0131\u015fmas\u0131 \u00f6nermektedir. Lee ve ark. (2019), u\u00e7tan uca e\u011fitilmi\u015f bir uzamsal transformat\u00f6r a\u011f\u0131 arac\u0131l\u0131\u011f\u0131yla bir \u015fekil \u015fablonunun ilgilenilen temel yap\u0131ya uyacak \u015fekilde deforme edildi\u011fi \u015fablon transformat\u00f6r a\u011flar\u0131n\u0131 benimsemi\u015ftir. <\/p>\n\n\n\n<p>     Yang ve di\u011ferleri. (2019), iki b\u00f6l\u00fcmden olu\u015fan ay\u0131rt edici bir koroner arter izleme y\u00f6ntemini benimsemi\u015ftir: bir izleyici ve bir ay\u0131r\u0131c\u0131. Fu ve ark. (2020), pulmoner damarlardan kaynaklanan m\u00fcdahaleleri \u00f6nlemek i\u00e7in akci\u011fer b\u00f6lgesinin \u00f6nceden maskelendi\u011fi koroner arter segmentasyonu i\u00e7in Mask R-CNN&#8217;yi kulland\u0131. Lei ve di\u011ferleri. (2020), CCTA g\u00f6r\u00fcnt\u00fclerinden \u00e7\u0131kar\u0131lan bilgilendirici anlamsal \u00f6zellikleri vurgulamak i\u00e7in derin dikkat stratejisini tam evri\u015fimli a\u011f (FCN) modeline (Long ve di\u011ferleri, 2015) entegre etti. <\/p>\n\n\n\n<p>     Gu ve ark. (2020), daha iyi koroner arter segmentasyonu i\u00e7in k\u00fcresel \u00f6zellikli bir g\u00f6m\u00fcl\u00fc a\u011f \u00f6nerdi. Gu ve Cai (2021), hem 2D hem de 3D CNN&#8217;lerin avantajlar\u0131n\u0131 koruyan iki a\u015famal\u0131 bir strateji \u00f6nerdi. Zhu ve di\u011ferleri. (2021), daha kesin s\u0131n\u0131rlar elde etmek i\u00e7in \u00e7oklu seviyelerin ve farkl\u0131 al\u0131c\u0131 alanlar\u0131n \u00f6zelliklerini ayr\u0131 ayr\u0131 birle\u015ftiren uzay-zamansal \u00f6zellik f\u00fczyon yap\u0131s\u0131na dayal\u0131 U \u015feklinde bir a\u011f \u00f6nermi\u015ftir. <\/p>\n\n\n\n<p>     Liang ve di\u011ferleri. (2021), kafa kar\u0131\u015ft\u0131r\u0131c\u0131 kategorileri ve benzer g\u00f6r\u00fcn\u00fcm \u00f6zelliklerine sahip hedefleri ay\u0131rt etmek i\u00e7in kanal dikkatini ve mek\u00e2nsal dikkati entegre eden geli\u015fmi\u015f bir U-net \u00f6nerdi. Tian ve di\u011ferleri. (2021), s\u0131n\u0131rl\u0131 GPU belle\u011fi sorununu \u00e7\u00f6zmek i\u00e7in derin \u00f6\u011frenme ve dijital g\u00f6r\u00fcnt\u00fc i\u015fleme algoritmalar\u0131n\u0131 birle\u015ftirdi. Cheung ve di\u011ferleri. (2021), aort ve koroner arterleri segmentlere ay\u0131rmak i\u00e7in tam otomatik iki boyutlu bir U-net modeli \u00f6nerdi.<\/p>\n\n\n\n<p><strong>Tablo 2<\/strong><\/p>\n\n\n\n<p>\u2022 Mevcut derin \u00f6\u011frenmeye dayal\u0131 \u00e7al\u0131\u015fmalar ve son on y\u0131lda koroner arter segmentasyonunun Zar skorlar\u0131 (% olarak). \ud835\udc4e(\ud835\udc4f)&#8217;da \ud835\udc4e genel miktard\u0131r ve \ud835\udc4f e\u011fitim miktar\u0131d\u0131r.<br>\u2022 Genel, y\u00f6ntemin \u00e7e\u015fitli uygulamalar i\u00e7in tasarland\u0131\u011f\u0131n\u0131 belirtirken, \u00f6zel y\u00f6ntemin koroner damar segmentasyonu i\u00e7in tasarland\u0131\u011f\u0131n\u0131 ve optimize edildi\u011fini belirtir.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"694\" height=\"825\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-2-jpg.webp\" alt=\"Mevcut derin \u00f6\u011frenmeye dayal\u0131 \u00e7al\u0131\u015fmalar ve son on y\u0131lda koroner arter segmentasyonunun Zar skorlar\u0131 \" class=\"wp-image-884\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-2-jpg.webp 694w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-2-252x300.webp 252w\" sizes=\"auto, (max-width: 694px) 100vw, 694px\" \/><figcaption class=\"wp-element-caption\">Mevcut derin \u00f6\u011frenmeye dayal\u0131 \u00e7al\u0131\u015fmalar ve son on y\u0131lda koroner arter segmentasyonunun Zar skorlar\u0131 <\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"490\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-2-1-1024x490.webp\" alt=\"Genel, y\u00f6ntemin \u00e7e\u015fitli uygulamalar i\u00e7in tasarland\u0131\u011f\u0131n\u0131 belirtirken, \u00f6zel y\u00f6ntemin koroner damar segmentasyonu i\u00e7in tasarland\u0131\u011f\u0131n\u0131 ve optimize edildi\u011fini belirtir\" class=\"wp-image-885\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-2-1-1024x490.webp 1024w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-2-1-300x143.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-2-1-768x367.webp 768w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-2-1-jpg.webp 1054w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">Genel, y\u00f6ntemin \u00e7e\u015fitli uygulamalar i\u00e7in tasarland\u0131\u011f\u0131n\u0131 belirtirken, \u00f6zel y\u00f6ntemin koroner damar segmentasyonu i\u00e7in tasarland\u0131\u011f\u0131n\u0131 ve optimize edildi\u011fini belirtir<\/figcaption><\/figure>\n\n\n\n<p>     Do\u011frudan b\u00f6l\u00fcmleme, giri\u015fin yerel ayr\u0131nt\u0131lar\u0131 iyi bir \u015fekilde elde edememesi nedeniyle genellikle alt \u00f6rneklenmi\u015f bir g\u00f6r\u00fcnt\u00fc gerektirdi\u011finden, yama tabanl\u0131 b\u00f6l\u00fcmleme pop\u00fcler hale gelir. Temel olarak bu y\u00f6ntemde, ilk \u00f6nce alt \u00f6rneklenmi\u015f g\u00f6r\u00fcnt\u00fc \u00fczerinde kaba bir b\u00f6l\u00fcmleme ger\u00e7ekle\u015ftirilir ve ard\u0131ndan yerel ayr\u0131nt\u0131lar\u0131n \u00e7\u0131kar\u0131lmas\u0131 i\u00e7in yama tarz\u0131nda bir iyile\u015ftirme uygulan\u0131r. Duan ve di\u011ferleri (2018), damar geli\u015ftirme ve segmentasyon i\u00e7in geleneksel Hessian damar temelli yakla\u015f\u0131mdan daha iyi performans g\u00f6steren ba\u011flama duyarl\u0131 bir 3D FCN \u00f6nerdi (Frangi ve di\u011ferleri, 1998). Chen ve di\u011ferleri (2018b), hangi voksellerin damar l\u00fcmenine ait oldu\u011funu belirlemek i\u00e7in e\u015fle\u015ftirilmi\u015f \u00e7ok \u00f6l\u00e7ekli 3D CNN&#8217;yi benimsedi. <\/p>\n\n\n\n<p>     Huang ve di\u011ferleri (2018), CTA g\u00f6r\u00fcnt\u00fclerini k\u00fc\u00e7\u00fck par\u00e7alara d\u00f6n\u00fc\u015ft\u00fcrd\u00fc ve ard\u0131ndan bunlar\u0131 i\u015flenmek \u00fczere 3 boyutlu bir U-net&#8217;e (\u00c7i\u00e7ek vd., 2016) g\u00f6nderdi. Chen, koroner arterlerin boru \u015feklindeki yap\u0131s\u0131n\u0131 vurgulamak i\u00e7in damar haritalar\u0131n\u0131 3D U-Net&#8217;in girdisine dahil ediyor (\u00c7i\u00e7ek ve ark., 2016). Mirunalini ve ark. (2019), 2 boyutlu dilimlerde koroner arterlerin varl\u0131\u011f\u0131n\u0131 belirlemek i\u00e7in CNN ve tekrarlayan sinir a\u011flar\u0131n\u0131 birle\u015ftirdi. Wang ve di\u011ferleri. (2021), voksel ve nokta bulutu tabanl\u0131 segmentasyon y\u00f6ntemlerini kabadan inceye bir \u00e7er\u00e7eveye dahil etti. Pan ve ark. (2021), dengesizlik sorununun \u00fcstesinden gelmek i\u00e7in odak kayb\u0131yla daha da optimize edilen 3D Yo\u011fun-U-Net&#8217;i \u00f6nerdi.<\/p>\n\n\n\n<p>     Son zamanlarda g\u00f6r\u00fcnt\u00fc verilerinin a\u011fa\u00e7 yap\u0131lar\u0131 ve grafik yap\u0131lar\u0131 gibi \u00f6zel veri yap\u0131lar\u0131na d\u00f6n\u00fc\u015ft\u00fcr\u00fclerek segmentasyona dahil edilmesi i\u00e7in baz\u0131 y\u00f6ntemler \u00f6nerilmi\u015ftir. Koroner arterlerin morfolojik yap\u0131s\u0131n\u0131n a\u011faca benzedi\u011fine ve genellikle sol koroner arter ve sa\u011f koroner a\u011faca b\u00f6l\u00fcnd\u00fc\u011f\u00fcne ve kan\u0131n aorttan koroner arterlere ve daha sonra bireysel dallara akt\u0131\u011f\u0131na dikkat edin. Wolterink ve ark. (2019), koroner arteri b\u00f6l\u00fcmlere ay\u0131ran boru \u015feklindeki bir y\u00fczey a\u011f\u0131ndaki k\u00f6\u015felerin uzaysal konumunu tahmin etmek i\u00e7in grafik evri\u015fimli a\u011flar\u0131 (GCN&#8217;ler) kulland\u0131. Kong ve ark. (2020), koroner arterin anatomik yap\u0131s\u0131n\u0131 \u00f6\u011frenmek i\u00e7in yeni bir a\u011fa\u00e7 yap\u0131l\u0131 evri\u015fimli kap\u0131l\u0131 tekrarlayan birim (ConvGRU) modeli \u00f6nermi\u015ftir.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"karsilastirma-ve-veri-kumeleri\">Kar\u015f\u0131la\u015ft\u0131rma ve Veri K\u00fcmeleri<\/h3>\n\n\n\n<p>     Tablo 1 ve Tablo 2&#8217;de bu konunun d\u00fcnya \u00e7ap\u0131ndaki yayg\u0131n ilgisini rahatl\u0131kla g\u00f6rebiliyoruz. \u00c7o\u011fu eserin daha \u00f6nceki eserlerle adil ve kapsaml\u0131 bir kar\u015f\u0131la\u015ft\u0131rma yapmad\u0131\u011f\u0131n\u0131 da g\u00f6rebiliriz. Geleneksel makine \u00f6\u011frenimi tabanl\u0131 yakla\u015f\u0131m i\u00e7in bir\u00e7ok \u00e7al\u0131\u015fma, yaln\u0131zca 42 CTA g\u00f6r\u00fcnt\u00fcs\u00fcne sahip halka a\u00e7\u0131k bir veri k\u00fcmesine (Kiri\u015fli ve di\u011ferleri, 2013) dayal\u0131 algoritmik geli\u015ftirmeye odaklanm\u0131\u015ft\u0131r. Kar\u015f\u0131la\u015ft\u0131rmalar da bu veri setine dayal\u0131 olarak yap\u0131ld\u0131 ve her \u00e7al\u0131\u015fma (Mohr ve di\u011ferleri, 2012; Wang ve di\u011ferleri, 2012) yeniden uygulamaya gerek kalmadan di\u011fer makalelerin performans\u0131n\u0131 elde etti. <\/p>\n\n\n\n<p>     Bu arada, di\u011fer \u00e7al\u0131\u015fmalar (Han et al., 2016) de\u011ferlendirmeleri i\u00e7in ba\u015fka bir halka a\u00e7\u0131k veri k\u00fcmesini (Schaap et al., 2009a) kullan\u0131yordu. \u00d6te yandan, di\u011fer \u00e7al\u0131\u015fmalarla kar\u015f\u0131la\u015ft\u0131r\u0131lamayan veya yaln\u0131zca genel y\u00f6ntemlerle kar\u015f\u0131la\u015ft\u0131r\u0131lamayan, de\u011ferlendirme i\u00e7in \u015firket i\u00e7i veri setleri haz\u0131rlanan \u00e7ok daha fazla \u00e7al\u0131\u015fma var (Lugauer vd., 2014b; Chi vd., 2015; Han vd., 2016; Freiman ve di\u011ferleri, 2017; Gao ve di\u011ferleri, 2019; Du ve di\u011ferleri, 2021).<\/p>\n\n\n\n<p>     Burada genel y\u00f6ntemleri genellikle \u00e7e\u015fitli uygulamalar i\u00e7in genel algoritmalar\u0131 hedef alan y\u00f6ntemler olarak tan\u0131mlad\u0131\u011f\u0131m\u0131z\u0131, buna kar\u015f\u0131l\u0131k spesifik y\u00f6ntemleri ise \u00f6rne\u011fin koroner damar segmentasyonu gibi belirli uygulamalar i\u00e7in tasarlanm\u0131\u015f ve optimize edilmi\u015f y\u00f6ntemler olarak tan\u0131mlad\u0131\u011f\u0131m\u0131z\u0131 unutmay\u0131n. Tescilli veri setinin kullan\u0131m\u0131 ve dolay\u0131s\u0131yla eksik veya k\u0131smi kar\u015f\u0131la\u015ft\u0131rma sorunu, DL Tabanl\u0131 yakla\u015f\u0131m\u0131n h\u00e2kim olmaya ba\u015flamas\u0131yla daha da yayg\u0131n hale geldi. \u00d6rne\u011fin Kong et al. (2020), modelini daha \u00f6nceki ilgili \u00e7al\u0131\u015fmalarla kar\u015f\u0131la\u015ft\u0131rmad\u0131, yaln\u0131zca 3D U-net gibi baz\u0131 pop\u00fcler DNN&#8217;lerle kar\u015f\u0131la\u015ft\u0131rd\u0131 (Huang ve ark, 2019).<\/p>\n\n\n\n<p>     Shen ve di\u011ferleri. (2019), kendi \u00e7al\u0131\u015fmalar\u0131n\u0131n de\u011ferlendirilmesi i\u00e7in farkl\u0131 veri setlerindeki di\u011fer \u00e7al\u0131\u015fmalar\u0131n elde etti\u011fi Zar puanlar\u0131n\u0131 listeledi. Veri k\u00fcmesi tutars\u0131zl\u0131\u011f\u0131na ek olarak di\u011fer k\u0131yaslama bile\u015fenlerindeki sorunlar da yayg\u0131nd\u0131r. Baz\u0131 \u00e7al\u0131\u015fmalar (Chi ve ark, 2015; Han ve ark, 2016), spesifik klinik ihtiya\u00e7lar\u0131na g\u00f6re farkl\u0131 de\u011ferlendirme metrikleri setleri kullan\u0131rken, di\u011ferleri (Shen ve ark., 2019) ba\u015flang\u0131\u00e7taki aortun dahil edildi\u011fi farkl\u0131 bir a\u00e7\u0131klama stratejisi kulland\u0131 ve bu da \u00e7ok daha y\u00fcksek bir Zar puan\u0131na yol a\u00e7t\u0131.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"534\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-1-1024x534.webp\" alt=\"\u00d6nerilen ImageCAS veri k\u00fcmesindeki CT g\u00f6r\u00fcnt\u00fcleri, etiketleri, CT dilimleri ve CT dilimlerindeki etiketlerin dahil oldu\u011fu \u00f6rnekler\" class=\"wp-image-886\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-1-1024x534.webp 1024w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-1-300x156.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-1-768x400.webp 768w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-1-jpg.webp 1383w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u00d6nerilen ImageCAS veri k\u00fcmesindeki CT g\u00f6r\u00fcnt\u00fcleri, etiketleri, CT dilimleri ve CT dilimlerindeki etiketlerin dahil oldu\u011fu \u00f6rnekler<\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>\u015eekil 1:<\/strong> \u00d6nerilen ImageCAS veri k\u00fcmesindeki CT g\u00f6r\u00fcnt\u00fcleri, etiketleri, CT dilimleri ve CT dilimlerindeki etiketlerin dahil oldu\u011fu \u00f6rnekler. Koroner arterlerin alt s\u0131n\u0131flar\u0131n\u0131n ayr\u0131 ayr\u0131 etiketlenmedi\u011fini unutmay\u0131n.<\/p>\n\n\n\n<p>     Yukar\u0131daki analizle, bu konunun daha da geli\u015ftirilmesi i\u00e7in bir k\u0131yaslama ve veri setine ihtiya\u00e7 duyuldu\u011funu fark edebiliriz. Bir yandan, y\u00f6ntemlerin \u00e7o\u011fu de\u011ferlendirme i\u00e7in yaln\u0131zca \u015firket i\u00e7i veri k\u00fcmelerini kullan\u0131yor ancak g\u00f6r\u00fcnt\u00fc elde etme parametreleri, yeniden yap\u0131land\u0131rma teknikleri ve etiketleme y\u00f6ntemleri \u00f6nemli \u00f6l\u00e7\u00fcde farkl\u0131l\u0131k g\u00f6steriyor. \u00d6te yandan, y\u00f6ntemlerin \u00e7o\u011fu genel y\u00f6ntemlerle kar\u015f\u0131la\u015ft\u0131rmalar yap\u0131yordu ancak ilgili \u00e7al\u0131\u015fmalar yoktu. Derin \u00f6\u011frenmeye dayal\u0131 y\u00f6ntemlerde ise durum daha da k\u00f6t\u00fcle\u015fti. <\/p>\n\n\n\n<p>     Derin \u00f6\u011frenmenin ba\u015far\u0131s\u0131n\u0131n b\u00fcy\u00fck \u00f6l\u00e7\u00fcde y\u00fcksek kaliteli a\u00e7\u0131klamalara sahip b\u00fcy\u00fck bir veri k\u00fcmesine ba\u011fl\u0131 oldu\u011funu ancak yayg\u0131n olarak kullan\u0131lan iki veri k\u00fcmesinin olduk\u00e7a k\u00fc\u00e7\u00fck oldu\u011funu (e\u011fitim i\u00e7in s\u0131ras\u0131yla yaln\u0131zca 8 ve 18 g\u00f6r\u00fcnt\u00fc i\u00e7erir) unutmay\u0131n. Ayr\u0131ca hemen hemen t\u00fcm eserlerin kaynak kodlar\u0131n\u0131n verilmemesi, adil bir kar\u015f\u0131la\u015ft\u0131rma yap\u0131lmas\u0131 a\u00e7\u0131s\u0131ndan da bir\u00e7ok zorlu\u011fu beraberinde getirmektedir. <\/p>\n\n\n\n<p>     Her ne kadar baz\u0131 \u00e7al\u0131\u015fmalar\u0131n kar\u015f\u0131la\u015ft\u0131rma amac\u0131yla uygulanmas\u0131 nispeten kolay olsa da, di\u011ferlerinin \u00e7o\u011fu, a\u011f yap\u0131s\u0131, \u00f6n i\u015fleme, hiper parametreler ve son i\u015fleme gibi spesifikasyonlar\u0131nda \u00e7ok fazla ayr\u0131nt\u0131 i\u00e7eren yo\u011fun \u015fekilde karma\u015f\u0131kt\u0131r. Yukar\u0131daki sorun di\u011fer baz\u0131 \u00e7al\u0131\u015fmalarda da tespit edilmi\u015ftir. \u00d6rne\u011fin, \u00e7al\u0131\u015fma (Tian ve di\u011ferleri, 2021) &#8220;koroner arter segmentasyon y\u00f6ntemlerinin \u00e7o\u011fu \u00f6zel veri seti kulland\u0131\u011f\u0131ndan, test etmemiz i\u00e7in uygun bir genel koroner veri seti bulunmad\u0131\u011f\u0131n\u0131&#8221; belirtmi\u015ftir.<\/p>\n\n\n\n<p>     \u00d6rne\u011fin, (Tian et al., 2021) \u00e7al\u0131\u015fmas\u0131nda &#8220;koroner arter segmentasyon y\u00f6ntemlerinin \u00e7o\u011fu \u00f6zel veri seti kulland\u0131\u011f\u0131ndan, test etmemiz i\u00e7in uygun bir genel koroner veri seti bulunmad\u0131\u011f\u0131n\u0131&#8221; belirtti. Hem Huang ve ark. (2018) hem de Yang ve ark. (2019), veri seti \u00fczerinde ilgili \u00e7al\u0131\u015fmalarla bir kar\u015f\u0131la\u015ft\u0131rma yapm\u0131\u015f (Schaap vd., 2009b), ancak mevcut y\u00f6ntemleri uygulaman\u0131n kolay olmamas\u0131 nedeniyle \u00f6zel bir veri seti \u00fczerinde kar\u015f\u0131la\u015ft\u0131rma yapmadan denemeler yapm\u0131\u015ft\u0131r. <\/p>\n\n\n\n<p>     \u00c7al\u0131\u015fma (Huang ve di\u011ferleri, 2018) &#8220;e\u011fitim verilerinin nispeten k\u00fc\u00e7\u00fck g\u00f6r\u00fcnd\u00fc\u011f\u00fcn\u00fc&#8221; belirtti. Baz\u0131 \u00e7al\u0131\u015fmalar (Yang ve di\u011ferleri, 2019) da k\u00fc\u00e7\u00fck veri seti sorununu \u00e7\u00f6zmeye \u00e7al\u0131\u015ft\u0131. \u00c7al\u0131\u015fmada (Moeskops ve di\u011ferleri, 2016) &#8220;gelecekteki \u00e7al\u0131\u015fmalarda, mevcut mimarinin kapasitesini daha fazla veri ve segmentasyon g\u00f6revleriyle daha fazla ara\u015ft\u0131raca\u011f\u0131z&#8221; ifadesine yer verildi.<\/p>\n\n\n\n<p>     Her y\u00f6ntemin etkinli\u011fini ayn\u0131 \u015fekilde ara\u015ft\u0131rmak i\u00e7in b\u00fcy\u00fck bir veri seti toplad\u0131k ve bunu kamuya a\u00e7\u0131klad\u0131k. Bu veri seti, mevcut genel veri setlerinden olduk\u00e7a b\u00fcy\u00fck olan 1000 3D CTA g\u00f6r\u00fcnt\u00fcs\u00fc i\u00e7ermektedir. Ayr\u0131ca, mevcut tipik y\u00f6ntemleri uygulamak i\u00e7in elimizden gelenin en iyisini yapmaya \u00e7al\u0131\u015ft\u0131\u011f\u0131m\u0131z bir k\u0131yaslama \u00f6nerdik. Mevcut y\u00f6ntemlerin kodlar\u0131 kamuya a\u00e7\u0131klanmamas\u0131na ve baz\u0131 deneysel detaylar\u0131n eksik olmas\u0131na ra\u011fmen, onlar\u0131n temel fikirlerine dayanarak benzer uygulamalar\u0131 ger\u00e7ekle\u015ftirdik. Ayr\u0131ca, her y\u00f6ntemin performans\u0131n\u0131 birden fazla konfig\u00fcrasyon seti ile \u00f6nerilen veri seti \u00fczerinde de\u011ferlendirdik. Ayr\u0131ca mevcut \u00e7al\u0131\u015fmalardan daha iyi performans elde eden bir temel segmentasyon \u00e7er\u00e7evesi de \u00f6nerdik. \u00d6nerilen veri k\u00fcmesinin ve k\u0131yaslama noktas\u0131n\u0131n ayr\u0131nt\u0131lar\u0131 B\u00f6l\u00fcm 3 ve 4&#8217;te tart\u0131\u015f\u0131lacakt\u0131r.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"image-cad-veriseti\">ImageCAD Veriseti<\/h2>\n\n\n\n<p>     \u00d6nerilen veri seti, Siemens 128 kesitli \u00e7ift kaynakl\u0131 taray\u0131c\u0131 taraf\u0131ndan 1000 hastadan al\u0131nan 3D CTA g\u00f6r\u00fcnt\u00fclerinden olu\u015fmaktad\u0131r. Daha \u00f6nce koroner arter hastal\u0131\u011f\u0131 tan\u0131s\u0131 konmu\u015f hastalar i\u00e7in erken revask\u00fclarizasyon (sonraki 90 g\u00fcn i\u00e7inde) dahildir. Y\u00fcksek doz BTA yap\u0131l\u0131r ve rekonstr\u00fcksiyon s\u0131ras\u0131nda en iyi koroner arter g\u00f6r\u00fcnt\u00fclerini elde etmek i\u00e7in %30-40 veya %60-70 faz\u0131 se\u00e7ilir. G\u00f6r\u00fcnt\u00fcler 512 \u00d7 512 \u00d7 (206\u2212 275) voksel boyutlar\u0131na, 0,29\u20130,43 mm2 d\u00fczlemsel \u00e7\u00f6z\u00fcn\u00fcrl\u00fc\u011fe ve 0,25\u20130,45 mm aral\u0131\u011fa sahiptir. Veriler, Nisan 2012&#8217;den Aral\u0131k 2018&#8217;e kadar Guangdong Eyalet Halk Hastanesindeki ger\u00e7ek\u00e7i klinik vakalardan topland\u0131. <\/p>\n\n\n\n<p>     Yaln\u0131zca 18 ya\u015f\u0131ndan b\u00fcy\u00fck ve iskemik inme, ge\u00e7ici iskemik atak ve\/veya periferik arter hastal\u0131\u011f\u0131 ile ilgili t\u0131bbi ge\u00e7mi\u015fi belgelenmi\u015f olan hastalar dahil edilmeye uygundur. \u0130ndeks kardiyak BTA veya CCTA&#8217;n\u0131n d\u00fc\u015f\u00fck g\u00f6r\u00fcnt\u00fcleme kalitesi (seviye III radyolog taraf\u0131ndan de\u011ferlendirilen) de koroner arter fonksiyonu \u00fczerindeki olas\u0131 etkisi nedeniyle hari\u00e7 tutuldu. Son olarak toplam 414 kad\u0131n ve 586 erkek bulunmaktad\u0131r; ortalama ya\u015flar\u0131 s\u0131ras\u0131yla 59,98 ve 57,68&#8217;dir. Her g\u00f6r\u00fcnt\u00fcdeki sol ve sa\u011f koroner arterler iki radyolog taraf\u0131ndan ba\u011f\u0131ms\u0131z olarak etiketlenir ve sonu\u00e7lar\u0131 \u00e7apraz olarak do\u011frulan\u0131r. Tutars\u0131zl\u0131k durumunda \u00fc\u00e7\u00fcnc\u00fc bir radyolog a\u00e7\u0131klamay\u0131 yapacak ve nihai sonu\u00e7 fikir birli\u011fi ile belirlenecektir. <\/p>\n\n\n\n<p>     Etiketli koroner arter, sol ana koroner arteri i\u00e7erir, sol \u00f6n inen koroner arter, sol sirkumfleks koroner arter, sa\u011f koroner arter, diyagonal 1, diyagonal 2, \u00e7apraz 3, geni\u015f marjinal dal 1, geni\u015f marjinal dal 2, geni\u015f marjinal dal 3, ramus intermedius, arka inen arterler, AHA adland\u0131rma kural\u0131na g\u00f6re akut marjinal 1 ve di\u011fer kan damarlar\u0131 (17 paragraf). Veri k\u00fcmesinin iki \u00f6rne\u011fi \u015eekil 1&#8217;de g\u00f6sterilmektedir.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"karsilastirma\">Kar\u015f\u0131la\u015ft\u0131rma<\/h2>\n\n\n\n<p>     Son zamanlardaki neredeyse t\u00fcm \u00e7al\u0131\u015fmalarda derin \u00f6\u011frenme tabanl\u0131 y\u00f6ntemlerin koroner arter segmentasyonunda umut verici performans sergiledi\u011fi g\u00f6z \u00f6n\u00fcne al\u0131nd\u0131\u011f\u0131nda, do\u011frudan segmentasyon, yama tabanl\u0131 segmentasyon, a\u011fa\u00e7 veri tabanl\u0131 segmentasyon ve grafik veri tabanl\u0131 segmentasyon da dahil olmak \u00fczere birka\u00e7 tipik derin \u00f6\u011frenme tabanl\u0131 y\u00f6ntemi uygulad\u0131k. Ayr\u0131ca \u00f6nerilen b\u00fcy\u00fck \u00f6l\u00e7ekli veri seti \u00fczerinde mevcut \u00e7al\u0131\u015fmalardan daha iyi performans elde eden bir temel y\u00f6ntem de \u00f6neriyoruz. Her y\u00f6ntemin ayr\u0131nt\u0131lar\u0131n\u0131 a\u015fa\u011f\u0131daki b\u00f6l\u00fcmlerde a\u00e7\u0131klayaca\u011f\u0131z ve uygulamalar\u0131 hakk\u0131nda daha fazla bilgiyi <\/p>\n\n\n\n<p><a href=\"https:\/\/github.com\/XiaoweiXu\/ImageCAS-A-Large-Scale-Dataset-and- Benchmark-for-Coronary-Artery-Segmentation-based-on-CT\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/XiaoweiXu\/ImageCAS-A-Large-Scale-Dataset-and- Benchmark-for-Coronary-Artery-Segmentation-based-on-CT<\/a><\/p>\n\n\n\n<p>adresinde bulabilirsiniz.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"947\" height=\"957\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-2-jpg.webp\" alt=\"\u00d6nerilen k\u0131yaslamadaki y\u00f6ntemlere genel bak\u0131\u015f\" class=\"wp-image-887\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-2-jpg.webp 947w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-2-297x300.webp 297w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-2-768x776.webp 768w\" sizes=\"auto, (max-width: 947px) 100vw, 947px\" \/><figcaption class=\"wp-element-caption\">\u00d6nerilen k\u0131yaslamadaki y\u00f6ntemlere genel bak\u0131\u015f<\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>\u015eekil 2: <\/strong>\u00d6nerilen k\u0131yaslamadaki y\u00f6ntemlere genel bak\u0131\u015f: (a) do\u011frudan segmentasyon (Shen ve di\u011ferleri, 2019), (b) yama bazl\u0131 segmentasyon (Huang ve di\u011ferleri, 2018; Chen ve di\u011ferleri, 2019), (c) a\u011fa\u00e7 verilerine dayal\u0131 segmentasyon (Kong ve di\u011ferleri, 2020) ve (d) grafik verilerine dayal\u0131 segmentasyon (Wolterink ve di\u011ferleri, 2019).<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"dogrudan-segmentasyon\">Do\u011frudan Segmentasyon<\/h3>\n\n\n\n<p>     Do\u011frudan segmentasyon, yaln\u0131zca tek bir sinir a\u011f\u0131n\u0131 benimser; bu, koroner arter segmentasyonu i\u00e7in basit ama etkili bir y\u00f6ntemdir. \u00d6zellikle g\u00f6r\u00fcnt\u00fc, karma\u015f\u0131k i\u015flemler olmadan olas\u0131l\u0131k haritas\u0131n\u0131n \u00e7\u0131kt\u0131s\u0131n\u0131 veren a\u011fa beslenir. G\u00f6r\u00fcnt\u00fcler genellikle b\u00fcy\u00fck oldu\u011fundan ve dolay\u0131s\u0131yla GPU&#8217;lar taraf\u0131ndan s\u0131\u011fd\u0131r\u0131lmas\u0131 zor oldu\u011fundan, pratikte bunlar\u0131n daha k\u00fc\u00e7\u00fck bir \u00e7\u00f6z\u00fcn\u00fcrl\u00fc\u011fe k\u00fc\u00e7\u00fclt\u00fclmesi gerekir. Temsilci olarak FCN-AG&#8217;yi (Shen ve di\u011ferleri, 2019) se\u00e7iyoruz. Y\u00f6ntemin ayr\u0131nt\u0131lar\u0131 \u015eekil 2(A)&#8217;da g\u00f6sterilmektedir ve genel yap\u0131, dikkat kap\u0131s\u0131 mod\u00fcllerine sahip bir FCN omurgas\u0131d\u0131r. S\u00fcre\u00e7 a\u015fa\u011f\u0131daki \u00fc\u00e7 ad\u0131ma ayr\u0131labilir:<\/p>\n\n\n\n<p>(1) Enterpolasyon kullanarak y\u00fcksek \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc CTA g\u00f6r\u00fcnt\u00fclerini d\u00fc\u015f\u00fck \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc g\u00f6r\u00fcnt\u00fclere yeniden boyutland\u0131r\u0131n;<\/p>\n\n\n\n<p><br>(2) D\u00fc\u015f\u00fck \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc g\u00f6r\u00fcnt\u00fcy\u00fc FCN-AG a\u011f\u0131na besleyin ve ard\u0131ndan koroner arter tahminini elde edin.<\/p>\n\n\n\n<p><br>(3) Enterpolasyon kullanarak d\u00fc\u015f\u00fck \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc tahmin etiketlerini y\u00fcksek \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc etiketlere yeniden boyutland\u0131r\u0131n.<\/p>\n\n\n\n<p>Y\u00f6ntemin daha fazla ayr\u0131nt\u0131s\u0131 i\u00e7in l\u00fctfen orijinal \u00e7al\u0131\u015fmaya bak\u0131n (Shen ve di\u011ferleri, 2019).<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"yama-tabanli-segmentasyon\">Yama Tabanl\u0131 Segmentasyon<\/h2>\n\n\n\n<p>     Alt \u00f6rneklemeden kaynaklanan hesaplamal\u0131 kaynak k\u0131s\u0131tlamalar\u0131n\u0131n ve eksik ayr\u0131nt\u0131lar\u0131n \u00fcstesinden gelmek i\u00e7in yama tabanl\u0131 b\u00f6l\u00fcmleme (Huang ve di\u011ferleri, 2018; Chen ve di\u011ferleri, 2019) \u00f6nerilmi\u015ftir. Bu y\u00f6ntem s\u0131n\u0131f\u0131n\u0131 de\u011ferlendirmek i\u00e7in ikili CNN tabanl\u0131 \u00e7er\u00e7eveyi (Huang ve di\u011ferleri, 2018; Chen ve di\u011ferleri, 2019) \u00f6rnek olarak g\u00f6r\u00fcyoruz (\u015eekil 2(b)). \u0130lk olarak, ilgilenilen b\u00f6lgeyi (RoI) \u00e7\u0131karmak ve ilgisiz alanlar\u0131 kald\u0131rmak i\u00e7in bir 3D U-net kullan\u0131l\u0131r. <\/p>\n\n\n\n<p>     Ard\u0131ndan, kesme g\u00f6r\u00fcnt\u00fcs\u00fc 512 \u00d7 512 \u00d7 \ud835\udc4d&#8217;den \ud835\udc4a \u00d7 \ud835\udc3b \u00d7 \ud835\udc4d\ud835\udc50 (\ud835\udc4a \u2264 512,\ud835\udc3b \u2264 512,\ud835\udc4d \u2264 \ud835\udc4d\ud835\udc50 ) olarak yeniden boyutland\u0131r\u0131l\u0131r. (Chen ve ark., 2019)&#8217;a g\u00f6re, boru \u015feklindeki yap\u0131lar\u0131 geli\u015ftirmek i\u00e7in Frangi filtrelemesi kullan\u0131yoruz ve elde edilen vask\u00fcler iyile\u015ftirme haritas\u0131, \u00e7ok kanall\u0131 bir giri\u015f olu\u015fturmak i\u00e7in giri\u015f g\u00f6r\u00fcnt\u00fcs\u00fcyle birle\u015ftiriliyor. Daha sonra \u00e7ok kanall\u0131 giri\u015f, i\u015flenmek \u00fczere ba\u015fka bir CNN&#8217;ye beslenen k\u00fc\u00e7\u00fck par\u00e7alara ayr\u0131\u015ft\u0131r\u0131l\u0131r. Son olarak, nihai segmentasyon sonucunu elde etmek i\u00e7in yamalar\u0131n segmentasyonu birle\u015ftirilir.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"agac-verilerine-dayali-segmentasyon\">A\u011fa\u00e7 Verilerine Dayal\u0131 Segmentasyon<\/h2>\n\n\n\n<p>     A\u011fa\u00e7 verilerine dayal\u0131 segmentasyon y\u00f6ntemi, koroner arterlerin morfolojik yap\u0131s\u0131n\u0131 dikkate ald\u0131\u011f\u0131ndan umut vericidir. A\u011fa\u00e7 evri\u015fimli tekrarlayan sinir a\u011f\u0131n\u0131 uygulamak i\u00e7in Kong ve ark (2020)&#8217;da sunulan \u00e7al\u0131\u015fmay\u0131 se\u00e7iyoruz ve ayr\u0131nt\u0131lar \u015eekil 2(c)&#8217;de g\u00f6sterilmektedir. \u0130lk olarak, \u00f6nceden b\u00f6l\u00fcmlere ayr\u0131lm\u0131\u015f etiketlerin iskeletle\u015ftirilmesiyle yakla\u015f\u0131k merkez \u00e7izgisi elde edilir. <\/p>\n\n\n\n<p>     A\u011fac\u0131 olu\u015fturma ad\u0131mlar\u0131n\u0131 basitle\u015ftirmek i\u00e7in \ud835\udc4d eksenindeki en b\u00fcy\u00fck koordinata sahip nokta k\u00f6k d\u00fc\u011f\u00fcm olarak al\u0131n\u0131r ve geri kalanlar yaprak d\u00fc\u011f\u00fcmler olarak kabul edilir. Daha sonra, her merkez \u00e7izgisi noktas\u0131, daha sonra d\u00fc\u011f\u00fcm \u00f6zelliklerini \u00e7\u0131karmak i\u00e7in kullan\u0131lan yamay\u0131 elde etmek i\u00e7in merkez olarak al\u0131n\u0131r. Daha sonra a\u011fa\u00e7taki ba\u011flant\u0131lar merkez \u00e7izgisi noktalar\u0131 aras\u0131ndaki kom\u015fuluk ili\u015fkisine g\u00f6re olu\u015fturulur. Son olarak a\u011fa\u00e7 yap\u0131s\u0131 verileri TreeConvGRU&#8217;ya girdi olarak kullan\u0131l\u0131r (Kong ve di\u011ferleri, 2020) ve tahmin edilen etiketler \u00e7\u0131kt\u0131 olarak elde edilir.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"grafik-tabanli-segmentasyon\">Grafik Tabanl\u0131 Segmentasyon<\/h2>\n\n\n\n<p>     Grafik tabanl\u0131 segmentasyon, a\u011fa\u00e7 veri tabanl\u0131 segmentasyona benzer \u015fekilde \u00e7al\u0131\u015f\u0131r. Wolterink ve ark (2019)&#8217;teki fikirden yola \u00e7\u0131karak \u015eekil 2(d)&#8217;de g\u00f6sterilene benzer bir \u015fema tasarlad\u0131k. \u0130lk olarak merkez \u00e7izgisi, a\u011fa\u00e7 veri tabanl\u0131 segmentasyondakiyle ayn\u0131 i\u015flem hatt\u0131yla elde edilir. Daha sonra grafik yap\u0131l\u0131 veriler \u00fcretilir. \u00d6zellikle, her merkez \u00e7izgisi noktas\u0131n\u0131n te\u011fetine dik olan birka\u00e7 \u0131\u015f\u0131n yay\u0131nlan\u0131rken, biti\u015fik \u0131\u015f\u0131nlar aras\u0131ndaki iki te\u011fet taraf\u0131ndan olu\u015fturulan a\u00e7\u0131 ayn\u0131 kal\u0131r. <\/p>\n\n\n\n<p>     I\u015f\u0131n\u0131n i\u00e7inden ge\u00e7ti\u011fi voksel blo\u011funa ba\u011fl\u0131 olarak, \u0131\u015f\u0131n\u0131n ba\u015flang\u0131\u00e7 noktas\u0131ndan te\u011fet \u00e7izgisi y\u00f6n\u00fcnde belirli bir ad\u0131m boyutunda d boyutlu bir \u00f6zellik olu\u015fturulur. Her \u0131\u015f\u0131n damar\u0131n kenar\u0131yla kesi\u015fir ve kesi\u015fme noktas\u0131 ile merkez \u00e7izgisi noktas\u0131 aras\u0131ndaki \u00d6klid mesafesi yar\u0131\u00e7ap olarak elde edilir. T\u00fcm merkez hatt\u0131 noktalar\u0131n\u0131n yar\u0131\u00e7aplar\u0131 tahmin edildikten sonra koroner arterler yeniden yap\u0131land\u0131r\u0131l\u0131r ve son segmentasyonu elde edebiliriz.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"temel-yontem\">Temel Y\u00f6ntem<\/h2>\n\n\n\n<p>     \u00d6nerilen temel y\u00f6ntem, \u015eekil 3&#8217;te g\u00f6sterildi\u011fi gibi yama b\u00f6l\u00fcmleme ve kaba b\u00f6l\u00fcmlendirmenin bir kombinasyonudur. B\u00f6yle bir kombinasyon, performans ve uygulama fizibilitesi aras\u0131nda bir dengedir. Bir yandan, t\u00fcm 3D g\u00f6r\u00fcnt\u00fcn\u00fcn orijinal \u00e7\u00f6z\u00fcn\u00fcrl\u00fckte do\u011frudan b\u00f6l\u00fcmlenmesi, b\u00fcy\u00fck bellek t\u00fcketimi nedeniyle m\u00fcmk\u00fcn de\u011fildir; yeniden boyutland\u0131r\u0131lan g\u00f6r\u00fcnt\u00fcn\u00fcn b\u00f6l\u00fcmlenmesi (kaba b\u00f6l\u00fcmleme olarak tan\u0131mlan\u0131r) m\u00fcmk\u00fcnd\u00fcr ancak s\u0131n\u0131rl\u0131 bir performansa yol a\u00e7ar. \u00d6te yandan, yama tabanl\u0131 b\u00f6l\u00fcmlendirme daha fazla ayr\u0131nt\u0131 sa\u011flayabilir ancak bazen k\u00fcresel ba\u011flamsal bilgilerin kayb\u0131 nedeniyle bariz hatalara da neden olabilir.<\/p>\n\n\n\n<p>     Temel y\u00f6ntemimizin iki ana mod\u00fcl\u00fc vard\u0131r: yama b\u00f6l\u00fcmleme ve kaba b\u00f6l\u00fcmleme. \u0130ki mod\u00fcl taraf\u0131ndan i\u015flenmeden \u00f6nce, her giri\u015f g\u00f6r\u00fcnt\u00fcs\u00fc do\u011frusal enterpolasyon kullan\u0131larak 512 \u00d7 512 \u00d7 (206\u2212275)&#8217;ten 128 \u00d7 128 \u00d7 64&#8217;e yeniden boyutland\u0131r\u0131l\u0131r. Kaba segmentasyonda, genellikle global yap\u0131y\u0131 yakalayabilen koroner arterin kaba bir segmentasyonunu elde etmek i\u00e7in 3 boyutlu bir U-net (\u00c7i\u00e7ek ve ark., 2016) kullan\u0131l\u0131r; \u00f6rne\u011fin ilgili t\u00fcm damarlar hatal\u0131 s\u0131n\u0131rlarla dahil edilmi\u015ftir. Tart\u0131\u015fma kolayl\u0131\u011f\u0131 i\u00e7in, \ud835\udf03 (\ud835\udc4b)&#8217;yi segmentasyon a\u011f\u0131 olarak belirtiyoruz; burada \ud835\udc4b giri\u015f g\u00f6r\u00fcnt\u00fcs\u00fc ve \ud835\udc4c\u0302\ud835\udc50 ise \u00e7\u0131kt\u0131d\u0131r. Daha sonra a\u011f a\u015fa\u011f\u0131daki benzerlik katsay\u0131s\u0131 kayb\u0131yla e\u011fitilir:<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"990\" height=\"278\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/formul-1-jpg.webp\" alt=\"benzerlik katsay\u0131s\u0131 \" class=\"wp-image-890\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/formul-1-jpg.webp 990w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/formul-1-300x84.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/formul-1-768x216.webp 768w\" sizes=\"auto, (max-width: 990px) 100vw, 990px\" \/><figcaption class=\"wp-element-caption\">benzerlik katsay\u0131s\u0131 <\/figcaption><\/figure>\n\n\n\n<p>     Burada \ud835\udc4c temel ger\u00e7ektir ve \ud835\udc60 yumu\u015fatma fakt\u00f6r\u00fcd\u00fcr. Yama segmentasyonunda as\u0131l sorun, koroner arterleri i\u00e7eren ilgili yamalar\u0131n tam olarak nas\u0131l elde edilece\u011fidir. Yamalar olu\u015fturmak i\u00e7in basit\u00e7e kayan bir pencere kullan\u0131rsak, di\u011fer k\u00fc\u00e7\u00fck damarlar\u0131 veya damar benzeri anatomileri i\u00e7erenler de dahil olmak \u00fczere t\u00fcm yamalar i\u015flenecektir. Bu \u015fekilde, segmentasyon a\u011f\u0131n\u0131n hem hedef b\u00f6lgeleri hem de hedef olmayan b\u00f6lgeleri dikkate almas\u0131 gerekir, bu da segmentasyon g\u00f6revinde b\u00fcy\u00fck zorluk yarat\u0131r. Bu sorunu \u00e7\u00f6zmek i\u00e7in, genel hedef b\u00f6lgeyi kabaca elde etmek amac\u0131yla ba\u015fka bir a\u011f benimsiyoruz. <\/p>\n\n\n\n<p>     Kaba b\u00f6l\u00fcmleme ad\u0131m\u0131ndaki sonucun bu ama\u00e7 i\u00e7in benimsenebilece\u011fini unutmay\u0131n. Ancak \u00f6nceki deneylerde bunun damarlar\u0131n kesilmesi gibi \u00e7e\u015fitli sorunlara yol a\u00e7abilece\u011fini bulduk. Bu nedenle, 3D U-net kullanarak koroner arterin kaba bir maskesini \u00e7\u0131karmak i\u00e7in ba\u015fka bir alt ad\u0131m olan dilate damar segmentasyonunu sunuyoruz. Daha sonra \u00e7\u0131k\u0131\u015f geni\u015fler ve damarlar \u00e7ok daha kal\u0131n hale gelir ve kesintiye u\u011frama olas\u0131l\u0131\u011f\u0131 azal\u0131r. Damarlar\u0131 daha da kal\u0131n hale getirmek i\u00e7in a\u015fa\u011f\u0131daki a\u011f\u0131rl\u0131kl\u0131 benzerlik katsay\u0131s\u0131 kay\u0131p fonksiyonu kullan\u0131l\u0131r (Sudre vd., 2017):<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"407\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/formul-2-1024x407.webp\" alt=\"a\u011f\u0131rl\u0131kl\u0131 benzerlik katsay\u0131s\u0131 kay\u0131p fonksiyonu \" class=\"wp-image-891\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/formul-2-1024x407.webp 1024w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/formul-2-300x119.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/formul-2-768x305.webp 768w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/formul-2-jpg.webp 1305w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">a\u011f\u0131rl\u0131kl\u0131 benzerlik katsay\u0131s\u0131 kay\u0131p fonksiyonu <\/figcaption><\/figure>\n\n\n\n<p>     Burada \ud835\udf19(\ud835\udc4b) segmentasyon a\u011f\u0131, \ud835\udc4c\u0302\ud835\udc51 \u00e7\u0131kt\u0131, \ud835\udc60 yumu\u015fatma fakt\u00f6r\u00fc ve \ud835\udefc s\u0131n\u0131f a\u011f\u0131rl\u0131\u011f\u0131d\u0131r (\ud835\udefc \u2208 (0, 1)). Segmentasyon sonu\u00e7lar\u0131n\u0131n hedef b\u00f6lgelerin \u00e7o\u011funu kapsamas\u0131n\u0131 sa\u011flamak i\u00e7in \ud835\udefc 0,01 olarak ayarlanm\u0131\u015ft\u0131r. Bu nedenle a\u011f, \u00e7\u0131kt\u0131da b\u00fcy\u00fck boyutlu kaplar elde etmeye e\u011filimlidir. <\/p>\n\n\n\n<p>     A\u011f e\u011fitiminde temel olarak koroner arterin geni\u015fletilmesiyle \ud835\udc4c\u0302\ud835\udc51 elde edilir \ud835\udc4c. Kesin konumu elde etmek i\u00e7in \ud835\udc4c\u0302\ud835\udc51, enterpolasyon kullan\u0131larak giri\u015f g\u00f6r\u00fcnt\u00fcs\u00fcn\u00fcn orijinal boyutuna yeniden boyutland\u0131r\u0131l\u0131r. Test a\u015famas\u0131nda geni\u015flemi\u015f damar segmentasyon a\u011f\u0131ndan elde edilen sonu\u00e7lar\u0131n daha iyi ba\u011flant\u0131 sa\u011flamak i\u00e7in daha da geni\u015fletildi\u011fini unutmay\u0131n. Damarlar\u0131n iskeleti, bir y\u00fczey inceltme algoritmas\u0131 kullan\u0131larak \u00e7\u0131kar\u0131l\u0131r (Lee ve di\u011ferleri, 1994).<\/p>\n\n\n\n<p>     Yeniden boyutland\u0131r\u0131lan iskelet ve giri\u015f g\u00f6r\u00fcnt\u00fcs\u00fcyle a\u015fa\u011f\u0131daki gibi bir yama segmentasyonu ger\u00e7ekle\u015ftiriyoruz:<\/p>\n\n\n\n<p>(1) Ba\u011flant\u0131l\u0131 bile\u015fen analizi yoluyla en b\u00fcy\u00fck iki ba\u011flant\u0131l\u0131 bile\u015fenin \u00e7\u0131kar\u0131lmas\u0131 ve di\u011ferlerinin at\u0131lmas\u0131. Bunun nedeni genellikle iki damar\u0131n (sol ve sa\u011f koroner arterler) bulundu\u011funa dair alan bilgisinden kaynaklanmaktad\u0131r;<\/p>\n\n\n\n<p><br>(2) Merkezinde iskelet noktas\u0131 ve kenar uzunlu\u011fu \ud835\udc5f olan \ud835\udc5b\ud835\udc50 k\u00fcbik par\u00e7a setlerinin \u00e7\u0131kar\u0131lmas\u0131 (\u015eekil 3&#8217;te \ud835\udc5b\ud835\udc50 = 3 ve \ud835\udc5f = 16, 32, 64);<\/p>\n\n\n\n<p><br>(3) 3D U-net++ (Zhou ve di\u011ferleri,2018), ayr\u0131nt\u0131lar\u0131 3D U-net&#8217;ten daha do\u011fru bir \u015fekilde i\u015fledi\u011finden \ud835\udc5b\ud835\udc50 yama setlerini i\u015flemek i\u00e7in benimsenmi\u015ftir;<\/p>\n\n\n\n<p><br>(4) Daha sonra b\u00f6l\u00fcmlere ayr\u0131lm\u0131\u015f yamalar, orijinal girdi g\u00f6r\u00fcnt\u00fcs\u00fcyle ayn\u0131 boyutta bir b\u00f6l\u00fcmleme g\u00f6r\u00fcnt\u00fcs\u00fc elde etmek \u00fczere birle\u015ftirilir;<\/p>\n\n\n\n<p>(5) Son olarak, \ud835\udc5b\ud835\udc50 segmentasyon g\u00f6r\u00fcnt\u00fcleri, son \u00e7\u0131kt\u0131y\u0131 elde etmek i\u00e7in topluluk ad\u0131m\u0131na beslenir. Burada topluluk i\u00e7in \u00e7o\u011funluk oylamas\u0131n\u0131n benimsendi\u011fini unutmay\u0131n.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"deneyler-ve-tartisma\">Deneyler ve Tart\u0131\u015fma<\/h2>\n\n\n\n<p>     Bu b\u00f6l\u00fcmde \u00f6ncelikle t\u00fcm deneylerin genel kurulumunu tart\u0131\u015f\u0131yoruz. Daha sonra her bir y\u00f6ntemin kendine \u00f6zg\u00fc konfig\u00fcrasyonlar\u0131 ve performanslar\u0131 tart\u0131\u015f\u0131lm\u0131\u015ft\u0131r. Son olarak, optimal konfig\u00fcrasyonlara sahip t\u00fcm y\u00f6ntemler kar\u015f\u0131la\u015ft\u0131r\u0131l\u0131r ve analiz edilir.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"523\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-3-1024x523.webp\" alt=\"\u00d6nerilen temel y\u00f6ntemin \u00e7er\u00e7evesi\" class=\"wp-image-892\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-3-1024x523.webp 1024w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-3-300x153.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-3-768x392.webp 768w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-3-jpg.webp 1408w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u00d6nerilen temel y\u00f6ntemin \u00e7er\u00e7evesi<\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>\u015eekil 3:<\/strong> \u00d6nerilen temel y\u00f6ntemin \u00e7er\u00e7evesi.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"deneme-kurulumu\">Deneme Kurulumu<\/h2>\n\n\n\n<p>     T\u00fcm deneyler PyTorch (Paszke ve di\u011ferleri, 2019) ve DGL (Wang ve di\u011ferleri, 2019) kullan\u0131larak ger\u00e7ekle\u015ftirildi ve 24G belle\u011fe sahip bir Nvidia RTX 3090 GPU \u00fczerinde ger\u00e7ekle\u015ftirildi. Do\u011frudan segmentasyon, yama segmentasyonu ve a\u011fa\u00e7 yap\u0131s\u0131 segmentasyonunda e\u011fitim s\u0131ras\u0131nda Zar kayb\u0131n\u0131 kulland\u0131k. Grafik tabanl\u0131 segmentasyonda, Wolterink&#8217;teki gibi kay\u0131p fonksiyonunda mesafe de\u011ferleri k\u00fcp olarak al\u0131n\u0131r ve e\u011fitim i\u00e7in Denklem (1) ve Denklem (2)&#8217;de g\u00f6sterilen kay\u0131p fonksiyonlar\u0131n\u0131 benimsedik. <\/p>\n\n\n\n<p>     Yama tabanl\u0131 segmentasyonda 16,32,64&#8217;l\u00fc yamalar tart\u0131\u015f\u0131l\u0131r ve e\u011fitim i\u00e7in Zar kayb\u0131 kullan\u0131l\u0131r. Baseline y\u00f6nteminde damarlar\u0131 geni\u015fletmek i\u00e7in yar\u0131\u00e7ap\u0131 \ud835\udc5f=5 olan k\u00fcresel bir yap\u0131 kullan\u0131l\u0131r. Uygulamam\u0131zdaki t\u00fcm a\u011flar 30 d\u00f6nem (yakla\u015f\u0131k 21.000 yineleme) i\u00e7in e\u011fitilmi\u015ftir ve Adam optimizasyonu 0,002 \u00f6\u011frenme oran\u0131yla benimsenmi\u015ftir. S\u0131n\u0131rl\u0131 GPU belle\u011fi nedeniyle, farkl\u0131 giri\u015f boyutlar\u0131 i\u00e7in toplu i\u015f boyutu farkl\u0131l\u0131k g\u00f6sterir. <\/p>\n\n\n\n<p>     \u00d6n b\u00f6l\u00fcmleme veya kaba b\u00f6l\u00fcmleme ad\u0131mlar\u0131nda 128 \u00d7 128 \u00d7 128, 256 \u00d7 256 \u00d7 128 ve 512 \u00d7 512 \u00d7 256 giri\u015f boyutu i\u00e7in toplu i\u015f boyutu s\u0131ras\u0131yla 8, 2 ve 1&#8217;dir. Yama segmentasyon ad\u0131m\u0131nda bu, 163, 323 ve 643 giri\u015f boyutu i\u00e7in s\u0131ras\u0131yla 512, 64, 8&#8217;dir. Deneyler, 750 vakadan olu\u015fan bir e\u011fitim seti (do\u011frulama i\u00e7in 50 vaka kullan\u0131l\u0131r) ve 250 vakadan olu\u015fan bir test seti ile 4 katl\u0131 \u00e7apraz do\u011frulama yakla\u015f\u0131m\u0131 kullan\u0131larak de\u011ferlendirildi. De\u011ferlendirme i\u00e7in, Tablo 1 ve Tablo 2&#8217;de belirtildi\u011fi gibi toplumda yayg\u0131n olarak kullan\u0131lan Zar puan\u0131 kullan\u0131lmaktad\u0131r.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"yapilandirma-tartismasi\">Yap\u0131land\u0131rma Tart\u0131\u015fmas\u0131<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"dogrudan-segmentasyon-1\">Do\u011frudan Segmentasyon<\/h3>\n\n\n\n<p>     Giri\u015f boyutu, dikkat mekanizmas\u0131n\u0131n kullan\u0131m\u0131 ve kanal say\u0131s\u0131 gibi segmentasyon performans\u0131n\u0131 etkileyebilecek baz\u0131 fakt\u00f6rleri tart\u0131\u015ft\u0131k. Orijinal g\u00f6r\u00fcnt\u00fcn\u00fcn en yak\u0131n kom\u015fu enterpolasyonuyla elde edilen 128 \u00d7 128 \u00d7 128, 256 \u00d7 256 \u00d7 128 ve 512 \u00d7 512 \u00d7 256&#8217;y\u0131 i\u00e7eren girdi boyutlar\u0131 tart\u0131\u015f\u0131lmaktad\u0131r. 4 ve 12 olmak \u00fczere kanal say\u0131s\u0131 tart\u0131\u015f\u0131lmaktad\u0131r.<\/p>\n\n\n\n<p>     Sonu\u00e7lar \u015eekil 4&#8217;te g\u00f6sterilmektedir. 512 \u00d7 512 \u00d7 256 giri\u015f boyutunun, s\u0131ras\u0131yla 256 \u00d7 256 \u00d7 128 ve 128 \u00d7 128 \u00d7 128&#8217;e g\u00f6re Zar skorunu \u00f6nemli \u00f6l\u00e7\u00fcde %7.38 (\ud835\udc5d &lt;0.0001) ve %12.32 (\ud835\udc5d &lt;0.0001) art\u0131rd\u0131\u011f\u0131n\u0131 g\u00f6zlemleyebiliriz. Dikkat kap\u0131 mod\u00fcl\u00fcn\u00fcn eklenmesi, performans\u0131 \u00f6nemli \u00f6l\u00e7\u00fcde %1.34 (\ud835\udc5d &lt;0.0001) art\u0131r\u0131r; bu durum ayn\u0131 zamanda Shen ve ark. (2019) \u00e7al\u0131\u015fmas\u0131nda da g\u00f6zlemlenmi\u015ftir. 5.24M parametre say\u0131s\u0131na sahip olan 12 kanal, 0.59M parametre say\u0131s\u0131na sahip olan 4 kanala g\u00f6re %2.13 (\ud835\udc5d &lt; 0.0001) daha y\u00fcksek bir Zar skoru elde eder.<\/p>\n\n\n\n<p>     G\u00f6rsel tart\u0131\u015fma \u015eekil 5&#8217;te g\u00f6sterilmektedir. BTA g\u00f6r\u00fcnt\u00fclerinde koroner b\u00f6lge etraf\u0131ndaki kontrast\u0131n d\u00fc\u015f\u00fck olmas\u0131 nedeniyle do\u011frudan segmentasyonda yerel detaylar g\u00f6z ard\u0131 edilerek her zaman koroner arterin tamam\u0131na odaklan\u0131l\u0131r. Vask\u00fcler k\u0131s\u0131m ile biti\u015fik dokunun geri kalan\u0131 aras\u0131nda voksel yo\u011funlu\u011fu a\u00e7\u0131s\u0131ndan \u00e7ok az fark oldu\u011funu ve do\u011frudan segmentasyon y\u00f6nteminin koroner arterin bu b\u00f6l\u00fcm\u00fcn\u00fc do\u011fru \u015fekilde tan\u0131mlayamad\u0131\u011f\u0131 ve bunun sonucunda segmentasyon hatalar\u0131na yol a\u00e7t\u0131\u011f\u0131n\u0131 fark edebiliriz. Ayr\u0131ca, y\u00fcksek \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc bir girdiyle do\u011frudan b\u00f6l\u00fcmleme daha fazla hesaplama kayna\u011f\u0131 gerektirir ve a\u011f boyutunu ve model kapasitesini s\u0131n\u0131rlar.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"yama-tabanli-segmentasyon-2\">Yama Tabanl\u0131 Segmentasyon<\/h3>\n\n\n\n<p>     Uygulamada, Chen&#8217;in kulland\u0131\u011f\u0131 bi\u00e7imde, e\u011fitim setinde ger\u00e7ek etiketlerin iskeletle\u015ftirilmesi i\u00e7in ger\u00e7ek etiketlerin kullan\u0131ld\u0131\u011f\u0131 bir k\u0131rpma y\u00f6ntemi benimsendi. Kar\u015f\u0131l\u0131k gelen yamalar iskelet noktalar\u0131n\u0131n ortas\u0131ndan k\u0131rp\u0131l\u0131r ve k\u0131rp\u0131lan b\u00f6lgeler, etiketli b\u00f6lgelerin olmayanlara oran\u0131 1:1 olacak \u015fekilde rastgele se\u00e7ilir. Performans\u0131 etkileyebilecek yama boyutu, frangi kanal\u0131 ve veri art\u0131rma dahil \u00fc\u00e7 fakt\u00f6r\u00fc ara\u015ft\u0131rd\u0131k. Veri b\u00fcy\u00fctme i\u00e7in d\u00f6nme olas\u0131l\u0131\u011f\u0131 (rastgele 0, 90, 180, 270 derece) ve yatay \u00e7evirme tart\u0131\u015f\u0131lmaktad\u0131r. Tart\u0131\u015fma kolayl\u0131\u011f\u0131 a\u00e7\u0131s\u0131ndan iki olas\u0131l\u0131k ayn\u0131 de\u011fere ayarlanm\u0131\u015ft\u0131r ve 0, 0,2 ve 0,5 dahil \u00fc\u00e7 de\u011fer tart\u0131\u015f\u0131lm\u0131\u015ft\u0131r.<\/p>\n\n\n\n<p>     Niceliksel performans \u015eekil 6&#8217;da g\u00f6sterilmektedir. Frangi kanall\u0131 a\u011f ile olmayan a\u011f aras\u0131nda Dice skorunda sadece %0,01 (p&gt;0,05) farkla anlaml\u0131 bir fark yoktur. Yama boyutu a\u00e7\u0131s\u0131ndan, daha b\u00fcy\u00fck bir yama boyutunun, daha k\u00fc\u00e7\u00fck bir yama boyutuna g\u00f6re \u00f6nemli \u00f6l\u00e7\u00fcde daha y\u00fcksek bir Zar puan\u0131 elde etti\u011fini fark edebiliriz; bu, daha b\u00fcy\u00fck bir yama boyutunun daha b\u00fcy\u00fck bir al\u0131c\u0131 alana sahip olmas\u0131 ve dolay\u0131s\u0131yla daha iyi ba\u011flam bilgisi yakalayabilmesi nedeniyle beklenen bir durumdur.<\/p>\n\n\n\n<p>     Veri art\u0131rma i\u00e7in, 0\/0 (d\u00f6nd\u00fcrme ve yans\u0131tma olmaks\u0131z\u0131n) flip ve rotasyon olas\u0131l\u0131klar\u0131n\u0131n, s\u0131ras\u0131yla 0.2\/0.2 ve 0.5\/0.5&#8217;e g\u00f6re Dice skorlar\u0131nda %2.63 (\ud835\udc5d &lt; 0.0001) ve %2.73 (\ud835\udc5d &lt; 0.0001) art\u0131\u015f elde etti\u011fini g\u00f6zlemliyoruz. Bu ilgin\u00e7 fenomen, koroner arterlerin \u00e7evredeki anatomilere kar\u015f\u0131l\u0131k gelen y\u00f6nlere sahip olmas\u0131 ve rotasyon ve ters \u00e7evirme operasyonlar\u0131n\u0131n, e\u011fitim s\u00fcrecine zarar verebilecek ger\u00e7ek\u00e7i olmayan art\u0131r\u0131lm\u0131\u015f \u00f6rnekler \u00fcretebilmesi olabilir.<\/p>\n\n\n\n<p>     Yama bazl\u0131 segmentasyonda ba\u015far\u0131s\u0131z k\u0131rp\u0131lm\u0131\u015f ilgi b\u00f6lgelerinin (RoI) g\u00f6rsel tart\u0131\u015fmas\u0131 \u015eekil 7&#8217;de g\u00f6sterilmektedir. G\u00f6r\u00fcn\u00fcmlerinin koroner arterlere olduk\u00e7a benzemesinden kaynaklanabilecek bir\u00e7ok benzer damar\u0131n koroner arter olarak tan\u0131nd\u0131\u011f\u0131n\u0131 g\u00f6rebiliriz. Bu sorunun \u00fcstesinden gelmek i\u00e7in, i\u015flem sonras\u0131 ad\u0131mda bir ba\u011flant\u0131 alan\u0131 analizini benimseyebiliriz ve tahmin edilen g\u00f6r\u00fcnt\u00fclerin \u00e7o\u011fu i\u00e7in bu iyile\u015ftirmenin, benzer b\u00f6lgelerin \u00e7\u0131kar\u0131lmas\u0131nda etkili oldu\u011funu bulduk. Ancak, bu baz\u0131 g\u00f6r\u00fcnt\u00fclerde iyi \u00e7al\u0131\u015fmaz; Fig\u00fcr 8&#8217;de g\u00f6sterildi\u011fi gibi, bir\u00e7ok k\u00fc\u00e7\u00fck damar\u0131n olu\u015fturulmas\u0131na ve koroner arterlerin k\u0131smi olarak kald\u0131r\u0131lmas\u0131na neden olur. Fig\u00fcr 8(a)&#8217;da s\u0131n\u0131r, kalp \u00e7evresindeki dokudan ba\u015far\u0131l\u0131 bir \u015fekilde kald\u0131r\u0131l\u0131rken, Fig\u00fcr 8(b)&#8217;de \u00f6n-segmentasyon, \u00e7\u0131k\u0131\u015fta kemik dokusu gibi di\u011fer yap\u0131lar\u0131 tutar.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"413\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-4-1024x413.webp\" alt=\"do\u011frudan segmentasyonun performans \" class=\"wp-image-893\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-4-1024x413.webp 1024w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-4-300x121.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-4-768x310.webp 768w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-4-jpg.webp 1215w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">do\u011frudan segmentasyonun performans <\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>\u015eekil 4:<\/strong> (a) giri\u015f boyutu, (b) dikkat kap\u0131s\u0131 mod\u00fcl\u00fc ve (c) kanal say\u0131s\u0131 dahil olmak \u00fczere \u00e7e\u015fitli konfig\u00fcrasyonlarla do\u011frudan segmentasyonun performans tart\u0131\u015fmas\u0131. ns anlaml\u0131 de\u011fil (\ud835\udc5d&gt;0,05) ve ** \ud835\udc5d 0,01&#8217;den k\u00fc\u00e7\u00fck, *** \ud835\udc5d 0,001&#8217;den k\u00fc\u00e7\u00fck ve **** \ud835\udc5d 0,0001&#8217;den k\u00fc\u00e7\u00fck anlam\u0131na gelir.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"826\" height=\"522\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-5-jpg.webp\" alt=\"do\u011frudan segmentasyon sonu\u00e7lar\u0131n\u0131n g\u00f6rsel tart\u0131\u015fmas\u0131\" class=\"wp-image-894\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-5-jpg.webp 826w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-5-300x190.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-5-768x485.webp 768w\" sizes=\"auto, (max-width: 826px) 100vw, 826px\" \/><figcaption class=\"wp-element-caption\">do\u011frudan segmentasyon sonu\u00e7lar\u0131n\u0131n g\u00f6rsel tart\u0131\u015fmas\u0131<\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>\u015eekil 5: <\/strong>Ar\u0131za durumlar\u0131yla do\u011frudan segmentasyon sonu\u00e7lar\u0131n\u0131n g\u00f6rsel tart\u0131\u015fmas\u0131<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"417\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-6-1024x417.webp\" alt=\"yama bazl\u0131 segmentasyonun performans tart\u0131\u015fmas\u0131\" class=\"wp-image-895\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-6-1024x417.webp 1024w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-6-300x122.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-6-768x313.webp 768w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-6-jpg.webp 1228w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">yama bazl\u0131 segmentasyonun performans tart\u0131\u015fmas\u0131<\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>\u015eekil 6:<\/strong> (a) Frangi kanal\u0131, (b) yama boyutu ve (c) veri art\u0131rma dahil olmak \u00fczere \u00e7e\u015fitli konfig\u00fcrasyonlarla yama bazl\u0131 segmentasyonun performans tart\u0131\u015fmas\u0131. ns anlaml\u0131 de\u011fil (\ud835\udc5d&gt;0,05) ve ** \ud835\udc5d 0,01&#8217;den k\u00fc\u00e7\u00fck, *** \ud835\udc5d 0,001&#8217;den k\u00fc\u00e7\u00fck ve **** \ud835\udc5d 0,0001&#8217;den k\u00fc\u00e7\u00fck anlam\u0131na gelir.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"489\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-7-1024x489.webp\" alt=\"yama tabanl\u0131 b\u00f6l\u00fcmlemede ba\u015far\u0131s\u0131z k\u0131rp\u0131lm\u0131\u015f yat\u0131r\u0131m getirisinin g\u00f6rseli\" class=\"wp-image-896\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-7-1024x489.webp 1024w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-7-300x143.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-7-768x367.webp 768w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-7-jpg.webp 1207w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">yama tabanl\u0131 b\u00f6l\u00fcmlemede ba\u015far\u0131s\u0131z k\u0131rp\u0131lm\u0131\u015f yat\u0131r\u0131m getirisinin g\u00f6rseli<\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>\u015eekil 7:<\/strong> Giri\u015f g\u00f6r\u00fcnt\u00fcleri, tahmin edilen etiketler, temel ger\u00e7ekler ve \u00f6nceden b\u00f6l\u00fcmlenmi\u015f olu\u015fturulmu\u015f s\u0131n\u0131rlay\u0131c\u0131 kutular \u00fczerinde kesitsel bir g\u00f6r\u00fcn\u00fcm de dahil olmak \u00fczere yama tabanl\u0131 b\u00f6l\u00fcmlemede ba\u015far\u0131s\u0131z k\u0131rp\u0131lm\u0131\u015f yat\u0131r\u0131m getirisinin g\u00f6rsel tart\u0131\u015fmas\u0131.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1023\" height=\"328\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-8-jpg.webp\" alt=\"Yama tabanl\u0131 segmentasyon\" class=\"wp-image-897\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-8-jpg.webp 1023w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-8-300x96.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-8-768x246.webp 768w\" sizes=\"auto, (max-width: 1023px) 100vw, 1023px\" \/><figcaption class=\"wp-element-caption\">Yama tabanl\u0131 segmentasyon<\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>\u015eekil 8:<\/strong> Yama tabanl\u0131 segmentasyonda en b\u00fcy\u00fck ba\u011flant\u0131l\u0131 bile\u015fen (LCC) son i\u015flemesinin (a) ba\u015far\u0131l\u0131 ve (b) ba\u015far\u0131s\u0131z durumlar\u0131n\u0131n g\u00f6rsel tart\u0131\u015fmas\u0131.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"351\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-9-1024x351.webp\" alt=\"A\u011fa\u00e7 veri tabanl\u0131 segmentasyon ve grafik tabanl\u0131 segmentasyon\" class=\"wp-image-898\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-9-1024x351.webp 1024w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-9-300x103.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-9-768x263.webp 768w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-9-jpg.webp 1214w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">A\u011fa\u00e7 veri tabanl\u0131 segmentasyon ve grafik tabanl\u0131 segmentasyon<\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>\u015eekil 9:<\/strong> A\u011fa\u00e7 veri tabanl\u0131 segmentasyon ve grafik tabanl\u0131 segmentasyonun performans tart\u0131\u015fmas\u0131, \u00e7e\u015fitli konfig\u00fcrasyonlar\u0131 i\u00e7eren (a) a\u011fa\u00e7 modelleri, (b) a\u011fa\u00e7 d\u00fc\u011f\u00fcm\u00fc olu\u015ftururken yama boyutu, (c) a\u011fa\u00e7 veri tabanl\u0131 segmentasyon i\u00e7in giri\u015f boyutu ve (d) grafik tabanl\u0131 segmentasyonu i\u00e7erir. ns, anlaml\u0131 de\u011fil anlam\u0131na gelir (\ud835\udc5d&gt;0.05), ve ** \ud835\udc5d&#8217;nin 0.01&#8217;den k\u00fc\u00e7\u00fck oldu\u011funu, *** \ud835\udc5d&#8217;nin 0.001&#8217;den k\u00fc\u00e7\u00fck oldu\u011funu, ve **** \ud835\udc5d&#8217;nin 0.0001&#8217;den k\u00fc\u00e7\u00fck oldu\u011funu ifade eder.<\/p>\n\n\n\n<p>     Ancak bu, baz\u0131 g\u00f6r\u00fcnt\u00fclerde pek i\u015fe yaramaz; \u00e7ok say\u0131da k\u00fc\u00e7\u00fck damar olu\u015fturur ve \u015eekil 8&#8217;de g\u00f6sterildi\u011fi gibi koroner arterlerin k\u0131smen \u00e7\u0131kar\u0131lmas\u0131yla sonu\u00e7lan\u0131r. \u015eekil 8(a)&#8217;da, s\u0131n\u0131r kalbin etraf\u0131ndaki dokudan iyice uzakla\u015ft\u0131r\u0131lm\u0131\u015fken, \u015eekil 8(b)&#8217;de \u00f6n segmentasyon, \u00e7\u0131kt\u0131da kemik dokusu gibi di\u011fer yap\u0131lar\u0131 tutar.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"agac-verisine-dayali-segmentasyon-ve-grafik-tabanli-segmentasyon\">A\u011fa\u00e7 Verisine Dayal\u0131 Segmentasyon ve Grafik Tabanl\u0131 Segmentasyon<\/h3>\n\n\n\n<p>     A\u011fa\u00e7 veri tabanl\u0131 segmentasyon ve grafik tabanl\u0131 segmentasyon benzer oldu\u011fundan ve ayn\u0131 \u00f6n segmentasyon mod\u00fcl\u00fcn\u00fc payla\u015ft\u0131klar\u0131ndan, tart\u0131\u015fma kolayl\u0131\u011f\u0131 i\u00e7in ikisini bir araya getirdik. Her iki y\u00f6ntemde de 128 \u00d7 128 \u00d7 128 ve 512 \u00d7 512 \u00d7 256&#8217;y\u0131 i\u00e7eren giri\u015f boyutu tart\u0131\u015f\u0131lmaktad\u0131r. A\u011fa\u00e7 yap\u0131s\u0131 segmentasyonunda, TreeConv3FGRU ve TreeConv3DLSTM dahil olmak \u00fczere iki modeli ve a\u011fa\u00e7 d\u00fc\u011f\u00fcm\u00fc in\u015fas\u0131 s\u0131ras\u0131nda tart\u0131\u015fma i\u00e7in 16 \u00d7 16 \u00d7 4 ve 16 \u00d7 16 \u00d7 8 i\u00e7eren yama boyutunu benimsedik. Segmentasyon performans\u0131 \u015eekil 9&#8217;da g\u00f6sterilmektedir. TreeConvGRU ve TreeConvLSTM adl\u0131 iki a\u011fa\u00e7 modelinin Zar skorunda yaln\u0131zca %0,06 (p&gt;0,05) farka sahip oldu\u011funu, dolay\u0131s\u0131yla \u00f6nemli bir fark olmad\u0131\u011f\u0131n\u0131 ke\u015ffedebiliriz.<\/p>\n\n\n\n<p>     A\u011fa\u00e7 verilerine dayal\u0131 segmentasyonda, 16 \u00d7 16 \u00d7 4&#8217;l\u00fck yama boyutu, 16 \u00d7 16 \u00d7 8&#8217;den %1,38&#8217;lik (\ud835\udc5d &lt; 0,005) daha y\u00fcksek bir Zar puan\u0131 elde eder; bu, daha b\u00fcy\u00fck yama boyutunun her zaman nihai performansa fayda sa\u011flayamayabilece\u011fini g\u00f6sterir. \u015eekil 9(c) ve (d) ile kar\u015f\u0131la\u015ft\u0131r\u0131ld\u0131\u011f\u0131nda, girdi boyutunun performans \u00fczerinde \u00e7ok farkl\u0131 bir etkiye sahip oldu\u011fu ilgin\u00e7 bir olguyu bulabiliriz. A\u011fa\u00e7 veri tabanl\u0131 segmentasyon i\u00e7in, y\u00fcksek \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc ve d\u00fc\u015f\u00fck \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc girdili uygulaman\u0131n performans\u0131 aras\u0131ndaki fark sadece %0,12&#8217;dir (p&gt;0,05). Grafik tabanl\u0131 segmentasyon i\u00e7in y\u00fcksek \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc girdiyle yap\u0131lan uygulama, Zar puan\u0131nda d\u00fc\u015f\u00fck \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc girdilerle yap\u0131lan uygulamadan %2,95 (\ud835\udc5d &lt; 0,0001) daha iyi performans g\u00f6steriyor.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"999\" height=\"318\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-10-jpg.webp\" alt=\"A\u011fa\u00e7 ve grafik tabanl\u0131 segmentasyon\" class=\"wp-image-899\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-10-jpg.webp 999w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-10-300x95.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-10-768x244.webp 768w\" sizes=\"auto, (max-width: 999px) 100vw, 999px\" \/><figcaption class=\"wp-element-caption\">A\u011fa\u00e7 ve grafik tabanl\u0131 segmentasyon<\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>\u015eekil 10:<\/strong> A\u011fa\u00e7 ve grafik tabanl\u0131 segmentasyonda ba\u015far\u0131s\u0131z vakalar\u0131n g\u00f6rsel tart\u0131\u015fmas\u0131.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"545\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-11-1024x545.webp\" alt=\"Temel y\u00f6ntem\" class=\"wp-image-900\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-11-1024x545.webp 1024w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-11-300x160.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-11-768x409.webp 768w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-11-jpg.webp 1030w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">Temel y\u00f6ntem<\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>\u015eekil 11: <\/strong>Temel y\u00f6ntemdeki (a) iyi ve (b) zay\u0131f b\u00f6l\u00fcmleme durumlar\u0131n\u0131n g\u00f6rsel tart\u0131\u015fmas\u0131.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"272\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-3-1024x272.webp\" alt=\"Temel y\u00f6ntemin zar puan\u0131\" class=\"wp-image-901\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-3-1024x272.webp 1024w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-3-300x80.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-3-768x204.webp 768w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-3-1536x408.webp 1536w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-3-jpg.webp 1557w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">Temel y\u00f6ntemin zar puan\u0131<\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>Tablo 3:<\/strong> Temel y\u00f6ntemin zar puan\u0131 (% olarak). Kaba b\u00f6l\u00fcmleme ad\u0131m\u0131ndan ve yama b\u00f6l\u00fcmlemesinin dallar\u0131ndan elde edilenleri i\u00e7eren ara sonu\u00e7lar da dahil edilmi\u015ftir.<\/p>\n\n\n\n<p>     Ayr\u0131ca, \u015eekil 10&#8217;da g\u00f6sterildi\u011fi gibi hem a\u011fa\u00e7 veri tabanl\u0131 b\u00f6l\u00fcmleme hem de grafik tabanl\u0131 b\u00f6l\u00fcmleme i\u00e7in kritik bir ad\u0131m olan \u00f6n b\u00f6l\u00fcmlemedeki ba\u015far\u0131s\u0131z durumlar\u0131 da tart\u0131\u015f\u0131yoruz. Her iki y\u00f6ntemin de merkez hatt\u0131 tabanl\u0131 \u00e7\u00f6z\u00fcmler oldu\u011funu ve verileri olu\u015fturmak ve damarlar\u0131 segmentlere ay\u0131rmak i\u00e7in merkez hatt\u0131na dayand\u0131\u011f\u0131n\u0131 unutmay\u0131n. <\/p>\n\n\n\n<p>     Pratik anlamda, bir merkez \u00e7izgisi modelinin e\u011fitimini destekleyecek ger\u00e7ek merkez \u00e7izgisi etiketlerinden yoksunuz. Burada, her iki y\u00f6ntemi de uygulamak i\u00e7in yakla\u015f\u0131k bir merkez \u00e7izgisi olu\u015fturmak \u00fczere daha sonra etiketlenen ve iskeletle\u015ftirilen a\u011f\u0131n b\u00f6l\u00fcmlendirilmesini benimsiyoruz. Bu nedenle, \u00f6n b\u00f6l\u00fcmlemenin kalitesi sonraki b\u00f6l\u00fcmlemenin do\u011frulu\u011funda \u00f6nemli bir fakt\u00f6r haline gelir. <\/p>\n\n\n\n<p>     \u015eekil 10&#8217;da g\u00f6sterildi\u011fi gibi \u00f6n segmentasyonda baz\u0131 koroner arterler eksiktir ve bunun sonucunda eksik damarlar bir a\u011fa\u00e7 yap\u0131s\u0131 veya grafik yap\u0131s\u0131 \u015feklinde olu\u015fturulmayacakt\u0131r. Son olarak, i\u015flemin geri kalan\u0131nda bu koroner arterler eksiktir ve bu da son segmentasyonda hatalara neden olur.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"temel-yontem-3\">Temel Y\u00f6ntem<\/h3>\n\n\n\n<p>     Mod\u00fcller (yama b\u00f6l\u00fcmlemesi ve kaba b\u00f6l\u00fcmleme), topluluk, geni\u015fleme (w\/o) ve yama boyutu (163, 323 ve 643) dahil olmak \u00fczere \u00e7e\u015fitli fakt\u00f6rleri tart\u0131\u015ft\u0131k. Performans Tablo 3&#8217;te g\u00f6sterilmektedir. Yama segmentasyon mod\u00fcl\u00fc i\u00e7in, s\u0131ras\u0131yla 163, 323 ve 643 boyutlar\u0131ndaki yamalar i\u00e7in dilatasyon olmadan elde edilen Zar skorlar\u0131 %79,56, %81,22 ve %82.34&#8217;t\u00fcr. \u0130kili kar\u015f\u0131la\u015ft\u0131rmalar\u0131 (163 ile 323 (\ud835\udc5d &lt;0.0001), 163 ile 643 (\ud835\udc5d &lt; 0.0001) ve 323 ile 643 (\ud835\udc5d &lt; 0.001)), istatistiksel olarak anlaml\u0131 farklar g\u00f6stermektedir. Daha b\u00fcy\u00fck yama boyutlar\u0131n\u0131n daha b\u00fcy\u00fck al\u0131c\u0131 alan\u0131 g\u00f6sterdi\u011fini ve bunun da segmentasyona fayda sa\u011flad\u0131\u011f\u0131n\u0131 fark edebiliriz. <\/p>\n\n\n\n<p>     Geni\u015fleme s\u0131ras\u0131nda, daha b\u00fcy\u00fck yama boyutlar\u0131 hala daha y\u00fcksek Zar puanlar\u0131 elde ediyor. Ancak 323 yama boyutu ile 643 yama boyutu aras\u0131nda istatistiksel bir anlaml\u0131l\u0131k (p&gt;0,05) yoktur. Bunun temel nedeni geni\u015flemenin ba\u011flam bilgisinin \u00e7\u0131kar\u0131lmas\u0131nda da etkili olmas\u0131d\u0131r. 323 yama boyutu ve geni\u015fletme kombinasyonu, ba\u011flam bilgisini \u00e7\u0131karmaya yetecek kadar g\u00fc\u00e7l\u00fcd\u00fcr ve daha b\u00fcy\u00fck yama boyutu ve geni\u015flemeye sahip daha b\u00fcy\u00fck al\u0131c\u0131 alan, daha fazla ba\u011flam bilgisi \u00e7\u0131karamaz. Ayr\u0131ca geni\u015fletmenin, 643 ve 323 yama boyutuyla performans\u0131 art\u0131rabildi\u011fini, ancak 163 i\u00e7in bu durumun ge\u00e7erli olmad\u0131\u011f\u0131n\u0131 da not edebiliriz; bunun nedeni, geni\u015fletmenin, giri\u015fte k\u00fc\u00e7\u00fck bir yama boyutuna sahip k\u00fc\u00e7\u00fck miktardaki baz\u0131 pikselleri g\u00f6z ard\u0131 etmesi olabilir.<\/p>\n\n\n\n<p>     Topluluk a\u00e7\u0131s\u0131ndan, kaba b\u00f6l\u00fcmleme mod\u00fcl\u00fcn\u00fc yama b\u00f6l\u00fcmleme mod\u00fcl\u00fcyle birlikte kullanan topluluk (%82,96), Zar puan\u0131 a\u00e7\u0131s\u0131ndan her bir temel s\u0131n\u0131fland\u0131r\u0131c\u0131dan (%77,80, %82,27, %82,70) daha iyi performans g\u00f6stermektedir. \u0130yi ve k\u00f6t\u00fc segmentasyon durumlar\u0131n\u0131n g\u00f6rsel g\u00f6sterimi \u015eekil 11&#8217;de g\u00f6sterilmektedir. \u015eekil 11(a)&#8217;da g\u00f6sterildi\u011fi gibi, segmentasyon sonucu iyidir ve hem 2D CT kesitlerinde hem de 3D g\u00f6r\u00fcn\u00fcmde temel ger\u00e7ekle e\u015fle\u015febilir.<\/p>\n\n\n\n<p>     Geni\u015fleme \u00e7\u0131kt\u0131s\u0131n\u0131n temel ger\u00e7e\u011fin t\u00fcm alanlar\u0131n\u0131 kapsad\u0131\u011f\u0131n\u0131 da g\u00f6rebiliriz. \u015eekil 11(b), zay\u0131f segmentasyonun bir \u00f6rne\u011fini g\u00f6stermektedir. \u015eekil 11(b), zay\u0131f segmentasyonun bir \u00f6rne\u011fini g\u00f6stermektedir. Kal\u0131n bir damar\u0131n ve uzun ince bir damar\u0131n eksik oldu\u011funu fark edebiliriz, bunun nedeni bunlar\u0131n dilate damar segmentasyon mod\u00fcl\u00fc taraf\u0131ndan tan\u0131nmamas\u0131d\u0131r. \u00d6zellikle 2D CT diliminde g\u00f6sterilen kal\u0131n damar sa\u011f atriyuma yak\u0131nd\u0131r ve sa\u011f atriyuma benzer gri tonlama de\u011ferine sahiptir. Sonu\u00e7 olarak damar\u0131n do\u011fru \u015fekilde tan\u0131nmas\u0131 nispeten zordur.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"735\" height=\"813\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-12-jpg.webp\" alt=\"Kar\u015f\u0131la\u015ft\u0131rmadaki be\u015f y\u00f6ntemin sonu\u00e7lar\u0131\" class=\"wp-image-902\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-12-jpg.webp 735w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/goruntu-12-271x300.webp 271w\" sizes=\"auto, (max-width: 735px) 100vw, 735px\" \/><figcaption class=\"wp-element-caption\">Kar\u015f\u0131la\u015ft\u0131rmadaki be\u015f y\u00f6ntemin sonu\u00e7lar\u0131<\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>\u015eekil 12: <\/strong>Kar\u015f\u0131la\u015ft\u0131rmadaki be\u015f y\u00f6ntemin sonu\u00e7lar\u0131n\u0131n d\u00f6rt vakayla g\u00f6rsel olarak kar\u015f\u0131la\u015ft\u0131r\u0131lmas\u0131<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"olcut-karsilastirmasi\">\u00d6l\u00e7\u00fct Kar\u015f\u0131la\u015ft\u0131rmas\u0131<\/h2>\n\n\n\n<p>     Kar\u015f\u0131la\u015ft\u0131rmadaki t\u00fcm y\u00f6ntemlerin performans kar\u015f\u0131la\u015ft\u0131rmas\u0131 Tablo 4&#8217;te g\u00f6sterilmektedir. \u00d6nerilen temel y\u00f6ntemin t\u00fcm \u00f6l\u00e7\u00fcmlerde optimum performansa ula\u015ft\u0131\u011f\u0131n\u0131 fark edebiliriz. Ayr\u0131ca yama b\u00f6l\u00fcmlemenin, a\u011fa\u00e7 veri tabanl\u0131 b\u00f6l\u00fcmlemenin ve grafik tabanl\u0131 b\u00f6l\u00fcmlendirmenin do\u011frudan b\u00f6l\u00fcmlendirmeden \u00e7ok daha d\u00fc\u015f\u00fck bir performansa sahip oldu\u011funu da ke\u015ffedebiliriz. Burada kaba bir tart\u0131\u015fma yapabiliriz.<\/p>\n\n\n\n<p>     Bu t\u00fcr bir olgu, do\u011frudan segmentasyonun (89 in Cheung et al. (2021)) yama segmentasyonu (94 in Pan et al. (2021)), a\u011fa\u00e7 veri tabanl\u0131 segmentasyon (85 in Kong et al. (2020)) ve Grafik tabanl\u0131 segmentasyondan daha y\u00fcksek Zar puan\u0131na sahip oldu\u011fu mevcut \u00e7al\u0131\u015fmalarda da ayn\u0131 e\u011filimi k\u0131smen g\u00f6stermektedir (73\u201375 in Wolterink et al. (2019)).<\/p>\n\n\n\n<p>     Ancak mevcut \u00e7al\u0131\u015fmalarda yama segmentasyonunun do\u011frudan segmentasyona g\u00f6re daha y\u00fcksek bir performansa sahip olmas\u0131 konusunda baz\u0131 tutars\u0131zl\u0131klar oldu\u011funu hala fark edebiliyoruz. Esas olarak iki sebep var.<\/p>\n\n\n\n<p>     Birincisi, de\u011ferlendirmeye y\u00f6nelik veri seti ayn\u0131 de\u011fildir ve veri setlerinin kalitesi de farkl\u0131l\u0131k g\u00f6stermektedir. Veri k\u00fcmemizin \u015fu anda en b\u00fcy\u00fc\u011f\u00fc oldu\u011funu ve mevcut y\u00f6ntemlerin kulland\u0131\u011f\u0131 \u00e7o\u011fu veri k\u00fcmesinden kat kat daha b\u00fcy\u00fck oldu\u011funu unutmay\u0131n.<\/p>\n\n\n\n<p>     \u0130kincisi, mevcut \u00e7al\u0131\u015fmalarda hiper parametreler, \u00f6n i\u015fleme ve son i\u015fleme gibi uygulamas\u0131 nispeten zor olan bir\u00e7ok detay bulunmaktad\u0131r.<\/p>\n\n\n\n<p>     Mevcut \u00e7al\u0131\u015fmalar\u0131 hayata ge\u00e7irmek i\u00e7in elimizden gelenin en iyisini yapmaya \u00e7al\u0131\u015fsak da anla\u015f\u0131lma eksikli\u011fi ve ilgili makalelerin uzunlu\u011funun s\u0131n\u0131rl\u0131 olmas\u0131 nedeniyle ka\u00e7\u0131n\u0131lmaz olarak g\u00f6zden ka\u00e7\u0131r\u0131lan baz\u0131 ayr\u0131nt\u0131lar (bazen kritik ayr\u0131nt\u0131lar) vard\u0131r. Bu nedenle, uygulamalar\u0131 geli\u015ftirmek i\u00e7in topluluktaki ilgili ara\u015ft\u0131rmac\u0131lar\u0131 aram\u0131za kat\u0131lmaya davet ediyoruz.<\/p>\n\n\n\n<p>     \u00d6nerilen k\u0131yaslamadaki y\u00f6ntemlerin g\u00f6rsel kar\u015f\u0131la\u015ft\u0131rmas\u0131 \u015eekil 12&#8217;de g\u00f6sterilmektedir. Do\u011frudan segmentasyon i\u00e7in Durum A ve Durum C&#8217;nin performans\u0131 iyidir. Ancak g\u00f6r\u00fcnt\u00fc kalitesi d\u00fc\u015f\u00fck oldu\u011funda ve durum B ve D&#8217;de g\u00f6sterildi\u011fi gibi koroner arterlerin yap\u0131s\u0131 \u00f6nemli \u00f6l\u00e7\u00fcde de\u011fi\u015fti\u011finde d\u00fc\u015f\u00fck kontrastl\u0131 alanlar do\u011fru \u015fekilde tespit edilemeyebilir.<\/p>\n\n\n\n<p>     Yama segmentasyonu i\u00e7in sonu\u00e7lar do\u011frudan segmentasyona olduk\u00e7a benzer. Tek fark, yama b\u00f6l\u00fctlemenin Durum D&#8217;deki baz\u0131 ince damarlar\u0131 tan\u0131yabilmesi ancak Durum A&#8217;da tan\u0131yamamas\u0131, do\u011frudan b\u00f6l\u00fctlemenin ise iki durumda damarlar \u00fczerinde z\u0131t bir performansa sahip olmas\u0131d\u0131r. Ayr\u0131ca yama b\u00f6l\u00fcmlendirmesi Durum B&#8217;de do\u011frudan b\u00f6l\u00fcmlemeye g\u00f6re biraz daha iyi bir performansa sahiptir.<\/p>\n\n\n\n<p>     G\u00f6r\u00fcn\u00fc\u015fe g\u00f6re yama b\u00f6l\u00fcmlendirmesi, yerel \u00f6zellikleri do\u011frudan b\u00f6l\u00fcmlendirme yamas\u0131ndan daha iyi i\u015fleyebilir; yama b\u00f6l\u00fcmlemenin yerel \u00f6zellik i\u015flemeye daha fazla dikkat etmesi bekleniyor. A\u011fa\u00e7 veri tabanl\u0131 b\u00f6l\u00fcmleme ve grafik tabanl\u0131 b\u00f6l\u00fcmleme olduk\u00e7a benzerdir ancak Durum B ve Durum D&#8217;de d\u00fc\u015f\u00fck kontrastl\u0131 damarlar\u0131 iyi bir \u015fekilde ke\u015ffedemezler.<\/p>\n\n\n\n<p>     Bunun temel nedeni, a\u011fa\u00e7 veri tabanl\u0131 ve grafik tabanl\u0131 segmentasyon y\u00f6ntemlerinin, a\u011fa\u00e7 ve grafik yap\u0131s\u0131ndaki d\u00fc\u011f\u00fcm say\u0131s\u0131n\u0131 belirleyen \u00f6n segmentasyon kullan\u0131larak \u00e7\u0131kar\u0131lan merkez \u00e7izgisine b\u00fcy\u00fck \u00f6l\u00e7\u00fcde dayanmas\u0131d\u0131r. \u00d6nerilen temel y\u00f6ntem i\u00e7in sonu\u00e7lar, \u00f6zellikle damarlar\u0131n \u00e7o\u011funlu\u011funun do\u011fru \u015fekilde tan\u0131nd\u0131\u011f\u0131 Durum B&#8217;de \u00e7ok daha iyi performans g\u00f6stermektedir. Bunun temel nedeni, temel y\u00f6ntemin hem kaba b\u00f6l\u00fcmleme hem de yama b\u00f6l\u00fcmlemesinden gelen \u00f6zellikleri \u00e7e\u015fitli yama boyutlar\u0131yla birle\u015ftirmesi ve b\u00f6ylece ba\u011flam bilgisini daha iyi \u00e7\u0131karabilmesidir.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"261\" src=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-4-1024x261.webp\" alt=\"Benchmarktaki y\u00f6ntemleri\" class=\"wp-image-903\" srcset=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-4-1024x261.webp 1024w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-4-300x76.webp 300w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-4-768x195.webp 768w, https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/tablo-4-jpg.webp 1352w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">Benchmarktaki y\u00f6ntemleri<\/figcaption><\/figure>\n\n\n\n<p class=\"has-text-align-center\"><strong>Tablo 4: <\/strong>Benchmarktaki y\u00f6ntemlerin Zar skorunda (%) performans kar\u015f\u0131la\u015ft\u0131rmas\u0131. Her y\u00f6ntem, \u00e7e\u015fitli konfig\u00fcrasyonlara sahip uygulamalar aras\u0131ndan se\u00e7ilen en uygun sonuca sahiptir.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"tartisma\">Tart\u0131\u015fma<\/h2>\n\n\n\n<p>     Do\u011frudan segmentasyon (Shen et al., 2019), yama tabanl\u0131 segmentasyon (Huang et al., 2018; Chen et al., 2019), a\u011fa\u00e7 verisine dayal\u0131 segmentasyon (Kong et al., 2020), grafik verisine dayal\u0131 segmentasyon (Wolterink et al., 2019) ve temel y\u00f6ntemimiz dahil olmak \u00fczere bir\u00e7ok tipik derin \u00f6\u011frenme tabanl\u0131 y\u00f6ntem uygulad\u0131k. <\/p>\n\n\n\n<p>     \u00d6nerilen temel y\u00f6ntem, Dice skoru, HD (Hausdorff uzakl\u0131\u011f\u0131) ve AHD (ortalama Hausdorff uzakl\u0131\u011f\u0131) \u00fczerinde en iyi performans\u0131 elde eder. \u00d6te yandan Kong et al. (2020).&#8217;un sonu\u00e7lar\u0131n\u0131n aksine a\u011fa\u00e7 veri tabanl\u0131 segmentasyon y\u00f6nteminin do\u011frudan segmentasyona g\u00f6re daha iyi bir performans elde etti\u011fini de g\u00f6rebiliriz. Bunun nedeni k\u0131smen Kong et al. (2020).&#8217;daki e\u011fitim ve optimizasyonda bir\u00e7ok teknik detay\u0131n bulunmas\u0131 olabilir. Adil bir kar\u015f\u0131la\u015ft\u0131rma i\u00e7in veri setimizi ve kodumuzu yay\u0131nlamam\u0131z\u0131n nedeni de budur.<\/p>\n\n\n\n<p>     Veri setimiz mevcut \u00e7al\u0131\u015fmalarla kar\u015f\u0131la\u015ft\u0131r\u0131ld\u0131\u011f\u0131nda olduk\u00e7a b\u00fcy\u00fck olmas\u0131na ve iki ila \u00fc\u00e7 deneyimli radyolog taraf\u0131ndan iyi bir \u015fekilde etiketlenmesine ra\u011fmen s\u0131n\u0131rlamalar\u0131 vard\u0131r. <\/p>\n\n\n\n<p>     \u00d6ncelikle veri setimiz tek merkezde toplanm\u0131\u015f ve dolay\u0131s\u0131yla yanl\u0131 da\u011f\u0131l\u0131mlara sahip olmu\u015ftur. \u0130kincisi, BT g\u00f6r\u00fcnt\u00fclerini elde etmek i\u00e7in yaln\u0131zca bir t\u00fcr CT makinesi, yani Siemens 128 kesitli \u00e7ift kaynakl\u0131 taray\u0131c\u0131 kullan\u0131ld\u0131, bu da \u00f6nyarg\u0131 sorununu daha da k\u00f6t\u00fcle\u015ftiriyor. \u00dc\u00e7\u00fcnc\u00fcs\u00fc, ayr\u0131nt\u0131l\u0131 etiketler sa\u011flanmam\u0131\u015ft\u0131r. <\/p>\n\n\n\n<p>     \u00d6rne\u011fin, sol ana koroner arter, sol \u00f6n inen koroner arter vb. dahil olmak \u00fczere koroner arterin alt s\u0131n\u0131flar\u0131 ayr\u0131lmam\u0131\u015ft\u0131r. Ba\u015fkalar\u0131n\u0131n da yukar\u0131daki s\u0131n\u0131rlamalar\u0131 hafifletmek ve ayn\u0131 zamanda ilgili ara\u015ft\u0131rmay\u0131 kolayla\u015ft\u0131rmak i\u00e7in veri k\u00fcmelerini yay\u0131nlayabileceklerini umuyoruz.<\/p>\n\n\n\n<p>     Kar\u015f\u0131la\u015ft\u0131rma \u00f6l\u00e7\u00fct\u00fcm\u00fcz\u00fcn gelecekteki y\u00f6nleri \u00e7e\u015fitli olabilir ve burada yaln\u0131zca birka\u00e7 tanesini isimlendirece\u011fiz.<\/p>\n\n\n\n<p>     \u0130lk olarak, performans\u0131 art\u0131rmak i\u00e7in nnU-net (Isensee ve di\u011ferleri, 2021), CoTr (Xie ve di\u011ferleri, 2021) ve UNETR (Hatamizadeh ve di\u011ferleri, 2022) gibi daha geli\u015fmi\u015f segmentasyon a\u011flar\u0131 mevcut \u00e7er\u00e7evelerde kullan\u0131labilir.<\/p>\n\n\n\n<p>     \u0130kincisi, daha ileri analiz ve tan\u0131da kritik bir rol oynayan koroner damarlar\u0131n topolojisini (Shit et al., 2021; Hu et al., 2019, 2021; Saeki et al., 2021) veya ba\u011flant\u0131s\u0131n\u0131 korumak i\u00e7in daha geli\u015fmi\u015f a\u011flar ve de\u011ferlendirme \u00f6l\u00e7\u00fcmleri ara\u015ft\u0131r\u0131labilir.<\/p>\n\n\n\n<p>     \u00dc\u00e7\u00fcnc\u00fcs\u00fc, veri setimize ve mevcut olanlara dayanarak (Schaap et al., 2009a; Kiri\u015fli et al., 2013), federal \u00f6\u011frenme (Rajasree et al., 2022) ve etki alan\u0131 uyarlamas\u0131 (Guan and Liu, 2021) dahil olmak \u00fczere \u00e7ok merkezli konular daha fazla ara\u015ft\u0131r\u0131labilir.<\/p>\n\n\n\n<h1 class=\"wp-block-heading\" id=\"cozum\">\u00c7\u00f6z\u00fcm<\/h1>\n\n\n\n<p>     Koroner arterlerin segmentasyonu, koroner arter hastal\u0131\u011f\u0131n\u0131n tan\u0131s\u0131 ve miktar\u0131n\u0131n belirlenmesi i\u00e7in kritik bir g\u00f6revdir. Bu yaz\u0131da, CTA g\u00f6r\u00fcnt\u00fclerinde koroner arter segmentasyonu i\u00e7in bir k\u0131yaslama veri seti \u00f6neriyoruz. Ek olarak, yaln\u0131zca mevcut birka\u00e7 tipik y\u00f6ntemi uygulamak i\u00e7in elimizden gelenin en iyisini yapmaya \u00e7al\u0131\u015fmad\u0131\u011f\u0131m\u0131z, ayn\u0131 zamanda g\u00fc\u00e7l\u00fc bir temel y\u00f6ntem \u00f6nerdi\u011fimiz bir k\u0131yaslama da uygulad\u0131k. <\/p>\n\n\n\n<p>     Kar\u015f\u0131la\u015ft\u0131rmal\u0131 de\u011ferlendirmedeki y\u00f6ntemlerin kapsaml\u0131 bir de\u011ferlendirmesini yapt\u0131k ve sonu\u00e7lar, \u00f6nerilen temel y\u00f6ntemin %82,96&#8217;l\u0131k bir Zar puan\u0131yla optimum performansa ula\u015ft\u0131\u011f\u0131n\u0131 g\u00f6steriyor.<\/p>\n\n\n\n<p>     Bununla birlikte, klinik uygulamalarda do\u011fru te\u015fhis ve stenoz miktar\u0131n\u0131n belirlenmesi i\u00e7in performans\u0131n hala geli\u015ftirilmeye ihtiyac\u0131 vard\u0131r. Kar\u015f\u0131la\u015ft\u0131rma ve veri k\u00fcmesi a\u015fa\u011f\u0131daki adreste yay\u0131nlanmaktad\u0131r.<\/p>\n\n\n\n<p><a href=\"https:\/\/github.com\/XiaoweiXu\/ImageCAS-A-Large-Scale-Dataset-and-Benchmark-for-Coronary-Artery-Segmentation-based-on-CT\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/XiaoweiXu\/ImageCAS-A-Large-Scale-Dataset-and-Benchmark-for-Coronary-Artery-Segmentation-based-on-CT<\/a><\/p>\n\n\n\n<p>     \u00d6nerilen veri k\u00fcmesinin ve kar\u015f\u0131la\u015ft\u0131rmal\u0131 de\u011ferlendirmenin toplulukta daha fazla ara\u015ft\u0131rmay\u0131 te\u015fvik edebilece\u011fini umuyoruz.<\/p>\n\n\n\n<h1 class=\"wp-block-heading\" id=\"yazar-katki-beyani\">Yazar Katk\u0131 Beyan\u0131<\/h1>\n\n\n\n<p><strong>An Zeng:<\/strong> Kavramsalla\u015ft\u0131rma, Metodoloji, Yaz\u0131l\u0131m.<br><strong>Chunbiao Wu:<\/strong> Veri toplama, Yazma \u2013 orijinal taslak.<br><strong>Guisen Lin:<\/strong> Kavramsalla\u015ft\u0131rma, Yazma \u2013 inceleme ve d\u00fczenleme.<br><strong>Wen Xie: <\/strong>Yazma \u2013 g\u00f6zden ge\u00e7irme &amp; d\u00fczenleme.<br><strong>Jin Hong:<\/strong> Yazma \u2013 inceleme ve d\u00fczenleme.<br><strong>Meiping Huang:<\/strong> Veri toplama, Ara\u015ft\u0131rma.<br><strong>Jian Zhuang: <\/strong>Veri toplama.<br><strong>Shanshan Bi: <\/strong>Veri toplama, Do\u011frulama.<br><strong>Dan Pan: <\/strong>Veri toplama, Yazma \u2013 inceleme ve d\u00fczenleme.<br><strong>Najeeb Ullah: <\/strong>Yazma \u2013 inceleme ve d\u00fczenleme.<br><strong>Kaleem Nawaz Khan: <\/strong>Yazma \u2013 inceleme ve d\u00fczenleme.<br><strong>Tianchen Wang: <\/strong>Veri toplama, Yazma \u2013 inceleme ve d\u00fczenleme.<br><strong>Yiyu Shi:<\/strong> Yazma \u2013 inceleme ve d\u00fczenleme.<br><strong>Xiaomeng Li:<\/strong> Kavramsalla\u015ft\u0131rma, Yazma \u2013 inceleme ve d\u00fczenleme.<br><strong>Xiaowei Xu:<\/strong> Kavramsalla\u015ft\u0131rma, Yazma \u2013 inceleme ve d\u00fczenleme, Denetim.<\/p>\n\n\n\n<h1 class=\"wp-block-heading\" id=\"rekabetci-menfaat-beyani\">Rekabet\u00e7i Menfaat Beyan\u0131<\/h1>\n\n\n\n<p>     Yazarlar, bu makalede rapor edilen \u00e7al\u0131\u015fmay\u0131 etkileyecek gibi g\u00f6r\u00fcnen, birbiriyle rekabet halinde olan herhangi bir finansal \u00e7\u0131kar veya ki\u015fisel ili\u015fkinin bulunmad\u0131\u011f\u0131n\u0131 beyan etmektedir.<\/p>\n\n\n\n<h1 class=\"wp-block-heading\" id=\"veri-kullanilabilirligi\">Veri Kullan\u0131labilirli\u011fi<\/h1>\n\n\n\n<p>Veri k\u00fcmesini bir ba\u011flant\u0131yla yay\u0131nlad\u0131k.<\/p>\n\n\n\n<h1 class=\"wp-block-heading\" id=\"tesekkurler\">Te\u015fekk\u00fcrler<\/h1>\n\n\n\n<p>     Bu \u00e7al\u0131\u015fma Guangdong Eyaleti Bilim ve Teknoloji Planlama Projesi taraf\u0131ndan desteklenmi\u015ftir. \u00c7in (No. 2019B020230003), Guangdong Zirve Projesi (No. DFJH201802), \u00c7in Ulusal Do\u011fa Bilimleri Vakf\u0131 (No. 62006050, No. 62276071), Guangzhou&#8217;daki Bilim ve Teknoloji Projeleri, \u00c7in (No. 202206010049, No. 2019A050510041), Guangdong Temel ve Uygulamal\u0131 Temel Ara\u015ft\u0131rma Vakf\u0131 (No. 2022A1515010157, 2022A1515011650), ve Guangzhou Bilim ve Teknoloji Planlama Projesi (No. 202102080188) ve Shenzhen&#8217;deki Sanming T\u0131p Projesi, \u00c7in (No. SZSM202011005) ve Sa\u011fl\u0131k Projesi, Guangdong Y\u00fcksek D\u00fczey Hastane \u0130n\u015faat\u0131 Fonu.<\/p>\n\n\n\n<h1 class=\"wp-block-heading\" id=\"etik-ve-bilgi-yonetimi-onaylari\">Etik ve Bilgi Y\u00f6netimi Onaylar\u0131<\/h1>\n\n\n\n<p>     Bu \u00e7al\u0131\u015fma ve z\u0131mni r\u0131zaya ili\u015fkin geriye d\u00f6n\u00fck verilerin toplanmas\u0131, 2019324H Protokol\u00fc kapsam\u0131nda Guangdong \u0130l Halk Hastanesi, Guangdong T\u0131p Bilimleri Akademisi&#8217;nden Ara\u015ft\u0131rma Etik Komitesi (REC) onay\u0131 ald\u0131. \u0130lgili t\u00fcm etik d\u00fczenlemelere uygundur. Kimlik tespiti, t\u00fcm CT dosyalar\u0131n\u0131n NIfTI format\u0131na d\u00f6n\u00fc\u015ft\u00fcr\u00fcld\u00fc\u011f\u00fc ve hastalar\u0131n ad\u0131, do\u011fum g\u00fcn\u00fc, kabul y\u0131l\u0131, kabul numaras\u0131 ve CT numaras\u0131 gibi hassas bilgilerinin kald\u0131r\u0131ld\u0131\u011f\u0131 bir s\u00fcre\u00e7te ger\u00e7ekle\u015ftirildi.<\/p>\n\n\n\n<h1 class=\"wp-block-heading\" id=\"referanslar\">Referanslar<\/h1>\n\n\n\n<p>Altunbay, D., Cigir, C., Sokmensuer, C., Gunduz-Demir, C., 2010. Color graphs for<br>automated cancer diagnosis and grading. IEEE Trans. Biomed. Eng. 57 (3),<br>665\u2013674.<\/p>\n\n\n\n<p>Aylward, S.R., Bullitt, E., 2002. Initialization, noise, singularities, and scale in height<br>ridge traversal for tubular object centerline extraction. IEEE Trans. Med. Imaging<br>21 (2), 61\u201375.<\/p>\n\n\n\n<p>Badrinarayanan, V., Kendall, A., Cipolla, R., 2017. Segnet: A deep convolutional<br>encoder-decoder architecture for image segmentation. IEEE Trans. Pattern Anal.<br>Mach. Intell. 39 (12), 2481\u20132495.<\/p>\n\n\n\n<p>Broersen, A., Kitslaar, P., Frenay, M., Dijkstra, J., 2012. FrenchCoast: fast, robust<br>extraction for the nice challenge on coronary artery segmentation of the tree. In:<br>Proc. of MICCAI Workshop\u2019\u2019 3D Cardiovascular Imaging: A MICCAI Segmentation<br>Challenge.<\/p>\n\n\n\n<p>Cetin, S., Unal, G., 2015. A higher-order tensor vessel tractography for segmentation<br>of vascular structures. IEEE Trans. Med. Imaging 34 (10), 2172\u20132185.<\/p>\n\n\n\n<p>Chen, H., Dou, Q., Yu, L., Qin, J., Heng, P.-A., 2018a. VoxResNet: Deep voxelwise<br>residual networks for brain segmentation from 3D MR images. NeuroImage 170,<br>446\u2013455.<\/p>\n\n\n\n<p>Chen, F., Li, Y., Tian, T., Cao, F., Liang, J., 2018b. Automatic coronary artery lumen<br>segmentation in computed tomography angiography using paired multi-scale 3D<br>CNN. In: Medical Imaging 2018: Biomedical Applications in Molecular, Structural,<br>and Functional Imaging, Vol. 10578. International Society for Optics and Photonics,<br>p. 105782R.<\/p>\n\n\n\n<p>Chen, Y.-C., Lin, Y.-C., Wang, C.-P., Lee, C.-Y., Lee, W.-J., Wang, T.-D., Chen, C.-<br>M., 2019. Coronary artery segmentation in cardiac CT angiography using 3D<br>multi-channel U-net. In: Medical Imaging with Deep Learning 2019.<\/p>\n\n\n\n<p>Cheung, W.K., Bell, R., Nair, A., Menezes, L.J., Patel, R., Wan, S., Chou, K., Chen, J.,<br>Torii, R., Davies, R.H., et al., 2021. A computationally efficient approach to<br>segmentation of the aorta and coronary arteries using deep learning. IEEE Access<br>9, 108873\u2013108888.<\/p>\n\n\n\n<p>Chi, Y., Huang, W., Zhou, J., Zhong, L., Tan, S.Y., Felix, K.Y.J., Sheon, L.C.S.,<br>San Tan, R., 2015. A composite of features for learning-based coronary artery<br>segmentation on cardiac CT angiography. In: International Workshop on Machine<br>Learning in Medical Imaging. Springer, pp. 271\u2013279.<\/p>\n\n\n\n<p>\u00c7i\u00e7ek, \u00d6., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O., 2016. 3D U-Net:<br>learning dense volumetric segmentation from sparse annotation. In: International<br>Conference on Medical Image Computing and Computer-Assisted Intervention.<br>Springer, pp. 424\u2013432.<\/p>\n\n\n\n<p>Collet, C., Onuma, Y., Andreini, D., Sonck, J., Pompilio, G., Mushtaq, S., La Meir, M.,<br>Miyazaki, Y., de Mey, J., Gaemperli, O., et al., 2018. Coronary computed tomography<br>angiography for heart team decision-making in multivessel coronary artery<br>disease. Eur. Heart J. 39 (41), 3689\u20133698.<\/p>\n\n\n\n<p>Cooper, R., Cutler, J., Desvigne-Nickens, P., Fortmann, S.P., Friedman, L., Havlik, R.,<br>Hogelin, G., Marler, J., McGovern, P., Morosco, G., et al., 2000. Trends and<br>disparities in coronary heart disease, stroke, and other cardiovascular diseases in<br>the United States: findings of the national conference on cardiovascular disease<br>prevention. Circulation 102 (25), 3137\u20133147.<\/p>\n\n\n\n<p>Diakogiannis, F.I., Waldner, F., Caccetta, P., Wu, C., 2020. ResUNet-a: A deep<br>learning framework for semantic segmentation of remotely sensed data. ISPRS J.<br>Photogramm. Remote Sens. 162, 94\u2013114.<\/p>\n\n\n\n<p>Douc, R., Moulines, E., Olsson, J., 2009. Optimality of the auxiliary particle filter.<br>Probab. Math. Statist. 29 (1), 1\u201328.<\/p>\n\n\n\n<p>Doyle, S., Madabhushi, A., Feldman, M., Tomaszeweski, J., 2006. A boosting cascade for<br>automated detection of prostate cancer from digitized histology. MICCAI 504\u2013511.<\/p>\n\n\n\n<p>Du, H., Shao, K., Bao, F., Zhang, Y., Gao, C., Wu, W., Zhang, C., 2021. Automated<br>coronary artery tree segmentation in coronary CTA using a multiobjective clustering<br>and toroidal model-guided tracking method. Comput. Methods Programs Biomed.<br>199, 105908.<\/p>\n\n\n\n<p>Duan, J., Bello, G., Schlemper, J., Bai, W., Dawes, T.J., Biffi, C., de Marvao, A.,<br>Doumoud, G., O\u2019Regan, D.P., Rueckert, D., 2019. Automatic 3D bi-ventricular<br>segmentation of cardiac images by a shape-refined multi-task deep learning<br>approach. IEEE Trans. Med. Imaging 38 (9), 2151\u20132164.<\/p>\n\n\n\n<p>Duan, Y., Feng, J., Lu, J., Zhou, J., 2018. Context aware 3D fully convolutional networks<br>for coronary artery segmentation. In: International Workshop on Statistical Atlases<br>and Computational Models of the Heart. Springer, pp. 85\u201393.<\/p>\n\n\n\n<p>Frangi, A.F., Niessen, W.J., Vincken, K.L., Viergever, M.A., 1998. Multiscale vessel<br>enhancement filtering. In: International Conference on Medical Image Computing<br>and Computer-Assisted Intervention. Springer, pp. 130\u2013137.<\/p>\n\n\n\n<p>Freiman, M., Nickisch, H., Prevrhal, S., Schmitt, H., Vembar, M., Maurovich-Horvat, P.,<br>Donnelly, P., Goshen, L., 2017. Improving CCTA-based lesions\u2019 hemodynamic<br>significance assessment by accounting for partial volume modeling in automatic<br>coronary lumen segmentation. Med. Phys. 44 (3), 1040\u20131049.<\/p>\n\n\n\n<p>Friman, O., Hindennach, M., K\u00fchnel, C., Peitgen, H.-O., 2010. Multiple hypothesis<br>template tracking of small 3D vessel structures. Med. Image Anal. 14 (2), 160\u2013171.<\/p>\n\n\n\n<p>Fu, Y., Guo, B., Lei, Y., Wang, T., Liu, T., Curran, W., Zhang, L., Yang, X., 2020. Mask<br>R-CNN based coronary artery segmentation in coronary computed tomography<br>angiography. In: Medical Imaging 2020: Computer-Aided Diagnosis, Vol. 11314.<br>International Society for Optics and Photonics, p. 113144F.<\/p>\n\n\n\n<p>Fu, H., Qiu, G., Shu, J., Ilyas, M., 2014. A novel polar space random field model for<br>the detection of glandular structures. IEEE Trans. Med. Imaging 33 (3), 764\u2013776.<\/p>\n\n\n\n<p>Gao, Z., Liu, X., Qi, S., Wu, W., Hau, W.K., Zhang, H., 2019. Automatic segmentation of<br>coronary tree in CT angiography images. Internat. J. Adapt. Control Signal Process.<br>33 (8), 1239\u20131247.<\/p>\n\n\n\n<p>Glover, G., Pelc, N., 1980. Nonlinear partial volume artifacts in x-ray computed<br>tomography. Med. Phys. 7 (3), 238\u2013248.<\/p>\n\n\n\n<p>Gu, L., Cai, X.-C., 2021. Fusing 2D and 3D convolutional neural networks for the<br>segmentation of aorta and coronary arteries from CT images. Artif. Intell. Med.<br>121, 102189.<\/p>\n\n\n\n<p>Gu, J., Fang, Z., Gao, Y., Tian, F., 2020. Segmentation of coronary arteries images using<br>global feature embedded network with active contour loss. Comput. Med. Imaging<br>Graph. 86, 101799.<\/p>\n\n\n\n<p>Guan, H., Liu, M., 2021. Domain adaptation for medical image analysis: a survey. IEEE<br>Trans. Biomed. Eng. 69 (3), 1173\u20131185.<\/p>\n\n\n\n<p>Gunduz-Demir, C., Kandemir, M., Tosun, A.B., Sokmensuer, C., 2010. Automatic<br>segmentation of colon glands using object-graphs. Med. Image Anal. 14 (1), 1\u201312.<\/p>\n\n\n\n<p>Han, D., Doan, N.-T., Shim, H., Jeon, B., Lee, H., Hong, Y., Chang, H.-J., 2014. A fast<br>seed detection using local geometrical feature for automatic tracking of coronary<br>arteries in CTA. Comput. Methods Programs Biomed. 117 (2), 179\u2013188.<\/p>\n\n\n\n<p>Han, D., Shim, H., Jeon, B., Jang, Y., Hong, Y., Jung, S., Ha, S., Chang, H.-J., 2016.<br>Automatic coronary artery segmentation using active search for branches and<br>seemingly disconnected vessel segments from coronary CT angiography. PLoS One<br>11 (8), e0156837.<\/p>\n\n\n\n<p>Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R.,<br>Xu, D., 2022. Unetr: Transformers for 3d medical image segmentation. In: Proceedings<br>of the IEEE\/CVF Winter Conference on Applications of Computer Vision. pp.<br>574\u2013584.<\/p>\n\n\n\n<p>Hu, X., Li, F., Samaras, D., Chen, C., 2019. Topology-preserving deep image<br>segmentation. Adv. Neural Inf. Process. Syst. 32.<\/p>\n\n\n\n<p>Hu, X., Wang, Y., Fuxin, L., Samaras, D., Chen, C., 2021. Topology-aware segmentation<br>using discrete morse theory. arXiv preprint arXiv:2103.09992.<\/p>\n\n\n\n<p>Huang, C., Han, H., Yao, Q., Zhu, S., Zhou, S.K., 2019. 3D U2-Net: A 3D universal<br>U-Net for multi-domain medical image segmentation. In: International Conference<br>on Medical Image Computing and Computer-Assisted Intervention. Springer, pp.<br>291\u2013299.<\/p>\n\n\n\n<p>Huang, W., Huang, L., Lin, Z., Huang, S., Chi, Y., Zhou, J., Zhang, J., Tan, R.-S.,<br>Zhong, L., 2018. Coronary artery segmentation by deep learning neural networks<br>on computed tomographic coronary angiographic images. In: 2018 40th Annual<br>International Conference of the IEEE Engineering in Medicine and Biology Society.<br>EMBC, IEEE, pp. 608\u2013611.<\/p>\n\n\n\n<p>Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H., 2021. nnU-Net: a<br>self-configuring method for deep learning-based biomedical image segmentation.<br>Nat. Methods 18 (2), 203\u2013211.<\/p>\n\n\n\n<p>Kiri\u015fli, H., Schaap, M., Metz, C., Dharampal, A., Meijboom, W.B., Papadopoulou, S.-<br>L., Dedic, A., Nieman, K., de Graaf, M.A., Meijs, M., et al., 2013. Standardized<br>evaluation framework for evaluating coronary artery stenosis detection, stenosis<br>quantification and lumen segmentation algorithms in computed tomography<br>angiography. Med. Image Anal. 17 (8), 859\u2013876.<\/p>\n\n\n\n<p>Kjerland, \u00d8., 2017. Segmentation of Coronary Arteries from CT-Scans of the Heart using<br>Deep Learning (Master\u2019s thesis). NTNU.<\/p>\n\n\n\n<p>Kong, B., Wang, X., Bai, J., Lu, Y., Gao, F., Cao, K., Xia, J., Song, Q., Yin, Y., 2020. Learning tree-structured representation for 3D coronary artery segmentation. Comput. Med. Imaging Graph. 80, 101688.<\/p>\n\n\n\n<p>Kroon, D.-J., 2009. Hessian Based Frangi Vesselness Filter. MATLAB Central File Exchange, https:\/\/www.mathworks.com\/matlabcentral\/fileexchange\/24409-hessianbased- frangi-vesselness-filter. <\/p>\n\n\n\n<p>Lee, T.-C., Kashyap, R.L., Chu, C.-N., 1994. Building skeleton models via 3-D medial surface axis thinning algorithms. CVGIP: Graph. Models Image Process. 56 (6), 462\u2013478. <\/p>\n\n\n\n<p>Lee, M.C.H., Petersen, K., Pawlowski, N., Glocker, B., Schaap, M., 2019. Template transformer networks for image segmentation. In: International Conference on Medical Imaging with Deep Learning\u2013Extended Abstract Track. <\/p>\n\n\n\n<p>Lei, Y., Guo, B., Fu, Y., Wang, T., Liu, T., Curran, W., Zhang, L., Yang, X., 2020. Automated coronary artery segmentation in coronary computed tomography angiography (CCTA) using deep learning neural networks. In: Medical Imaging 2020: Imaging Informatics for Healthcare, Research, and Applications, Vol. 11318. International Society for Optics and Photonics, 1131812. <\/p>\n\n\n\n<p>Lesage, D., Angelini, E.D., Funka-Lea, G., Bloch, I., 2016. Adaptive particle filtering for coronary artery segmentation from 3D CT angiograms. Comput. Vis. Image Underst. 151, 29\u201346. <\/p>\n\n\n\n<p>Li, X., Chen, H., Qi, X., Dou, Q., Fu, C.-W., Heng, P.-A., 2018a. H-DenseUNet: hybrid densely connected UNet for liver and tumor segmentation from CT volumes. IEEE Trans. Med. Imaging 37 (12), 2663\u20132674. <\/p>\n\n\n\n<p>Li, X., Dou, Q., Chen, H., Fu, C.-W., Qi, X., Belav`y, D.L., Armbrecht, G., Felsenberg, D., Zheng, G., Heng, P.-A., 2018b. 3D multi-scale FCN with random modality voxel dropout learning for intervertebral disc localization and segmentation from multi-modality MR images. Med. Image Anal. 45, 41\u201354. <\/p>\n\n\n\n<p>Li, Z., Zhang, X., M\u00fcller, H., Zhang, S., 2018c. Large-scale retrieval for medical image analytics: A comprehensive review. Med. Image Anal. 43, 66\u201384. <\/p>\n\n\n\n<p>Liang, R., Ma, J., Ma, G., Wang, K., et al., 2021. 3D U-Net with attention and focal loss for coronary tree segmentation. <\/p>\n\n\n\n<p>Lin, Y., Yao, H., Li, Z., Zheng, G., Li, X., 2022. Calibrating label distribution for class-imbalanced barely-supervised knee segmentation. arXiv preprint arXiv:2205. 03644. <\/p>\n\n\n\n<p>Long, J., Shelhamer, E., Darrell, T., 2015. Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3431\u20133440. <\/p>\n\n\n\n<p>Lugauer, F., Zhang, J., Zheng, Y., Hornegger, J., Kelm, B.M., 2014a. Improving accuracy in coronary lumen segmentation via explicit calcium exclusion, learning-based ray detection and surface optimization. In: Medical Imaging 2014: Image Processing, Vol. 9034. International Society for Optics and Photonics, p. 90343U. <\/p>\n\n\n\n<p>Lugauer, F., Zheng, Y., Hornegger, J., Kelm, B.M., 2014b. Precise lumen segmentation in coronary computed tomography angiography. In: International MICCAI Workshop on Medical Computer Vision. Springer, pp. 137\u2013147. <\/p>\n\n\n\n<p>Milletari, F., Navab, N., Ahmadi, S.-A., 2016. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 Fourth International Conference on 3D Vision (3DV). IEEE, pp. 565\u2013571. Mirunalini, P., Aravindan, C., Nambi, A.T., Poorvaja, S., Priya, V.P., 2019. Segmentation of coronary arteries from CTA axial slices using deep learning techniques. In: TENCON 2019-2019 IEEE Region 10 Conference. TENCON, IEEE, pp. 2074\u20132080. <\/p>\n\n\n\n<p>Moeskops, P., Wolterink, J.M., van der Velden, B.H., Gilhuijs, K.G., Leiner, T., Viergever, M.A., I\u0161gum, I., 2016. Deep learning for multi-task medical image segmentation in multiple modalities. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, pp. 478\u2013486. <\/p>\n\n\n\n<p>Mohr, B., Masood, S., Plakas, C., 2012. Accurate lumen segmentation and stenosis detection and quantification in coronary CTA. In: Proceedings of 3D Cardiovascular Imaging: A MICCAI Segmentation Challenge Workshop. <\/p>\n\n\n\n<p>Mortazi, A., Karim, R., Rhode, K., Burt, J., Bagci, U., 2017. CardiacNET: Segmentation of left atrium and proximal pulmonary veins from MRI using multi-view CNN. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, pp. 377\u2013385. <\/p>\n\n\n\n<p>Nguyen, K., Sarkar, A., Jain, A.K., 2012. Structure and context in prostatic gland segmentation and classification. In: MICCAI. Springer, pp. 115\u2013123.<\/p>\n\n\n\n<p>Nickisch, H., Lamash, Y., Prevrhal, S., Freiman, M., Vembar, M., Goshen, L., Schmitt, H., 2015. Learning patient-specific lumped models for interactive coronary blood flow simulations. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, pp. 433\u2013441. <\/p>\n\n\n\n<p>Oktay, O., Schlemper, J., Folgoc, L.L., Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N.Y., Kainz, B., et al., 2018. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999. <\/p>\n\n\n\n<p>Organization, W.H., et al., 2009. Cardiovascular diseases (cvds). http:\/\/www.who.int\/ mediacentre\/factsheets\/fs317\/en\/index.html. <\/p>\n\n\n\n<p>Pan, L.-S., Li, C.-W., Su, S.-F., Tay, S.-Y., Tran, Q.-V., Chan, W.P., 2021. Coronary artery segmentation under class imbalance using a U-Net based architecture on computed tomography angiography images. Sci. Rep. 11 (1), 1\u20137. <\/p>\n\n\n\n<p>Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al., 2019. Pytorch: An imperative style, high-performance deep learning library. Adv. Neural Inf. Process. Syst. 32. <\/p>\n\n\n\n<p>Patravali, J., Jain, S., Chilamkurthy, S., 2017. 2D-3D fully convolutional neural networks for cardiac MR segmentation. In: International Workshop on Statistical Atlases and Computational Models of the Heart. Springer, pp. 130\u2013139. <\/p>\n\n\n\n<p>Rajasree, R., Gopika, G., Romero, C.A.T., et al., 2022. The role and impact of federal learning in digital healthcare: A useful survey. In: Handbook of Research on Technical, Privacy, and Security Challenges in a Modern World. IGI Global, pp. 127\u2013147. <\/p>\n\n\n\n<p>Ronneberger, O., Fischer, P., Brox, T., 2015. U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, pp. 234\u2013241. <\/p>\n\n\n\n<p>Saeki, Y., Saito, A., Cousty, J., Kenmochi, Y., Shimizu, A., 2021. Statistical modeling of pulmonary vasculatures with topological priors in CT volumes. In: Interpretability of Machine Intelligence in Medical Image Computing, and Topological Data Analysis and its Applications for Medical Data: 4th International Workshop, iMIMIC 2021, and 1st International Workshop, TDA4MedicalData 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, September 27, 2021, Proceedings 4. Springer, pp. 108\u2013118. <\/p>\n\n\n\n<p>Schaap, M., Metz, C.T., van Walsum, T., van der Giessen, A.G., Weustink, A.C., Mollet, N.R., Bauer, C., Bogunovi\u0107, H., Castro, C., Deng, X., et al., 2009a. Standardized evaluation methodology and reference database for evaluating coronary artery centerline extraction algorithms. Med. Image Anal. 13 (5), 701\u2013714. <\/p>\n\n\n\n<p>Schaap, M., Metz, C., van Walsum, T., Niessen, W., 2009b. Rotterdam coronary artery algorithm evaluation framework. <\/p>\n\n\n\n<p>Shahzad, R., Kiri\u015fli, H., Metz, C., Tang, H., Schaap, M., van Vliet, L., Niessen, W., van Walsum, T., 2013. Automatic segmentation, detection and quantification of coronary artery stenoses on CTA. Int. J. Cardiovasc. Imaging 29 (8), 1847\u20131859.<\/p>\n\n\n\n<p>Shen, Y., Fang, Z., Gao, Y., Xiong, N., Zhong, C., Tang, X., 2019. Coronary arteries segmentation based on 3D FCN with attention gate and level set function. IEEE Access 7, 42826\u201342835. <\/p>\n\n\n\n<p>Shit, S., Paetzold, J.C., Sekuboyina, A., Ezhov, I., Unger, A., Zhylka, A., Pluim, J.P., Bauer, U., Menze, B.H., 2021. clDice-a novel topology-preserving loss function for tubular structure segmentation. In: Proceedings of the IEEE\/CVF Conference on Computer Vision and Pattern Recognition. pp. 16560\u201316569. <\/p>\n\n\n\n<p>Sirinukunwattana, K., Snead, D.R., Rajpoot, N.M., 2015a. A novel texture descriptor for detection of glandular structures in colon histology images. In: SPIE Medical Imaging. International Society for Optics and Photonics, p. 94200S. <\/p>\n\n\n\n<p>Sirinukunwattana, K., Snead, D.R., Rajpoot, N.M., 2015b. A stochastic polygons model for glandular structures in colon histology images. IEEE Trans. Med. Imaging 34 (11), 2366\u20132378. <\/p>\n\n\n\n<p>Skare, \u00d8., B\u00f8lviken, E., Holden, L., 2003. Improved sampling-importance resampling and reduced bias importance sampling. Scand. J. Stat. 30 (4), 719\u2013737. <\/p>\n\n\n\n<p>Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Cardoso, M.J., 2017. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support. Springer, pp. 240\u2013248. <\/p>\n\n\n\n<p>Tabesh, A., Teverovskiy, M., Pang, H.-Y., Kumar, V.P., Verbel, D., Kotsianti, A., Saidi, O., 2007. Multifeature prostate cancer diagnosis and Gleason grading of histological images. IEEE Trans. Med. Imaging 26 (10), 1366\u20131378. <\/p>\n\n\n\n<p>Tian, F., Gao, Y., Fang, Z., Gu, J., 2021. Automatic coronary artery segmentation algorithm based on deep learning and digital image processing. Appl. Intell. 51 (12), 8881\u20138895. <\/p>\n\n\n\n<p>Wang, C., Moreno, R., Smedby, \u00d6., 2012. Vessel segmentation using implicit modelguided level sets. In: MICCAI Workshop&#8221; 3D Cardiovascular Imaging: A MICCAI Segmentation Challenge&#8221;, Nice France, 1st of October 2012. <\/p>\n\n\n\n<p>Wang, M., Yu, L., Zheng, D., Gan, Q., Gai, Y., Ye, Z., Li, M., Zhou, J., Huang, Q., Ma, C., et al., 2019. Deep graph library: Towards efficient and scalable deep learning on graphs. <\/p>\n\n\n\n<p>Wang, Q., Zhao, W., Yan, X., Che, H., Ye, K., Lu, Y., Li, Z., Cui, S., 2021. Geometric morphology based irrelevant vessels removal for accurate coronary artery segmentation. In: 2021 IEEE 18th International Symposium on Biomedical Imaging. ISBI, IEEE, pp. 757\u2013760. <\/p>\n\n\n\n<p>Wolterink, J.M., Leiner, T., I\u0161gum, I., 2019. Graph convolutional networks for coronary artery segmentation in cardiac CT angiography. In: International Workshop on Graph Learning in Medical Imaging. Springer, pp. 62\u201369.<\/p>\n\n\n\n<p>Wu, D., Wang, X., Bai, J., Xu, X., Ouyang, B., Li, Y., Zhang, H., Song, Q., Cao, K.,<br>Yin, Y., 2019. Automated anatomical labeling of coronary arteries via bidirectional<br>tree LSTMs. Int. J. Comput. Assist. Radiol. Surg. 14 (2), 271\u2013280.<br><\/p>\n\n\n\n<p>Xie, Y., Zhang, J., Shen, C., Xia, Y., 2021. Cotr: Efficiently bridging cnn and transformer<br>for 3d medical image segmentation. In: Medical Image Computing and Computer<br>Assisted Intervention\u2013MICCAI 2021: 24th International Conference, Strasbourg,<br>France, September 27\u2013October 1, 2021, Proceedings, Part III 24. Springer, pp.<br>171\u2013180.<br><\/p>\n\n\n\n<p>Yang, H., Chen, J., Chi, Y., Xie, X., Hua, X., 2019. Discriminative coronary artery<br>tracking via 3d cnn in cardiac ct angiography. In: International Conference<br>on Medical Image Computing and Computer-Assisted Intervention. Springer, pp.<br>468\u2013476.<br><\/p>\n\n\n\n<p>Yu, L., Cheng, J.-Z., Dou, Q., Yang, X., Chen, H., Qin, J., Heng, P.-A., 2017. Automatic<br>3D cardiovascular MR segmentation with densely-connected volumetric convnets.<br>In: International Conference on Medical Image Computing and Computer-Assisted<br>Intervention. Springer, pp. 287\u2013295.<\/p>\n\n\n\n<p>Yu, L., Yang, X., Qin, J., Heng, P.-A., 2016. 3D FractalNet: dense volumetric segmentation for cardiovascular MRI volumes. In: Reconstruction, Segmentation, and Analysis of Medical Images. Springer, pp. 103\u2013110. <\/p>\n\n\n\n<p>Zhang, D.P., 2010. Coronary artery segmentation and motion modelling. <\/p>\n\n\n\n<p>Zhang, S., Fu, H., Yan, Y., Zhang, Y., Wu, Q., Yang, M., Tan, M., Xu, Y., 2019. Attention guided network for retinal image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, pp. 797\u2013805. <\/p>\n\n\n\n<p>Zheng, Y., Loziczonek, M., Georgescu, B., Zhou, S.K., Vega-Higuera, F., Comaniciu, D., 2011. Machine learning based vesselness measurement for coronary artery segmentation in cardiac CT volumes. In: Medical Imaging 2011: Image Processing, Vol. 7962. International Society for Optics and Photonics, p. 79621K. <\/p>\n\n\n\n<p>Zhou, Z., Siddiquee, M.M.R., Tajbakhsh, N., Liang, J., 2018. Unet++: A nested u-net architecture for medical image segmentation. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support. Springer, pp. 3\u201311. <\/p>\n\n\n\n<p>Zhu, H., Song, S., Xu, L., Song, A., Yang, B., 2021. Segmentation of coronary arteries images using spatio-temporal feature fusion network with combo loss. Cardiovasc. Eng. Technol. 1\u201312.<\/p>\n\n\n\n<p>Makalenin orijial halini a\u015fa\u011f\u0131da sizlerle payla\u015f\u0131yoruz.<\/p>\n\n\n\n<div data-wp-interactive=\"core\/file\" class=\"wp-block-file\"><object data-wp-bind--hidden=\"!state.hasPdfPreview\" hidden class=\"wp-block-file__embed\" data=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/zengCMIG2023.pdf\" type=\"application\/pdf\" style=\"width:100%;height:600px\" aria-label=\"zengCMIG2023 g\u00f6m\u00fcs\u00fc.\"><\/object><a id=\"wp-block-file--media-86962996-4fef-4b94-80f7-ccd40d77866c\" href=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/zengCMIG2023.pdf\">zengCMIG2023<\/a><a href=\"https:\/\/www.facadium.com.tr\/blog\/wp-content\/uploads\/2023\/12\/zengCMIG2023.pdf\" class=\"wp-block-file__button wp-element-button\" aria-describedby=\"wp-block-file--media-86962996-4fef-4b94-80f7-ccd40d77866c\" download>\u0130ndir<\/a><\/div>\n\n\n\n<p><\/p>\n\n\n\n<p><\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>IMAGECAS: B\u0130LG\u0130SAYARLI TOMOGRAF\u0130 ANJ\u0130YOGRAF\u0130 G\u00d6R\u00dcNT\u00dcLER\u0130NE DAYALI KORONER ARTER SEGMENTASYONU \u0130\u00c7\u0130N GEN\u0130\u015e \u00d6L\u00c7EKL\u0130 B\u0130R VER\u0130 SET\u0130 VE REFERANS NOKTASI \u00c7ALI\u015eMASI Not : Bilgisayarl\u0131 T\u0131bbi G\u00f6r\u00fcnt\u00fcleme Ve [&#8230;]<\/p>\n","protected":false},"author":3,"featured_media":905,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5,109],"tags":[32,31,34,8,30,9],"class_list":["post-882","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-python","category-yapay-zeka","tag-data-analysis","tag-data-mining","tag-data-science","tag-python","tag-veri-madenciligi","tag-yazilim"],"_links":{"self":[{"href":"https:\/\/www.facadium.com.tr\/blog\/wp-json\/wp\/v2\/posts\/882","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.facadium.com.tr\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.facadium.com.tr\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.facadium.com.tr\/blog\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/www.facadium.com.tr\/blog\/wp-json\/wp\/v2\/comments?post=882"}],"version-history":[{"count":8,"href":"https:\/\/www.facadium.com.tr\/blog\/wp-json\/wp\/v2\/posts\/882\/revisions"}],"predecessor-version":[{"id":914,"href":"https:\/\/www.facadium.com.tr\/blog\/wp-json\/wp\/v2\/posts\/882\/revisions\/914"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.facadium.com.tr\/blog\/wp-json\/wp\/v2\/media\/905"}],"wp:attachment":[{"href":"https:\/\/www.facadium.com.tr\/blog\/wp-json\/wp\/v2\/media?parent=882"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.facadium.com.tr\/blog\/wp-json\/wp\/v2\/categories?post=882"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.facadium.com.tr\/blog\/wp-json\/wp\/v2\/tags?post=882"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}